Glycemic indicators influence hypercoagulability via multisystem pathways in psychiatric inpatients: Structural equation modeling

Abstract

BACKGROUND

Dysglycemia and hypercoagulability frequently coexist in psychiatric inpatients, elevating venous thromboembolism risks. However, the specific biological cascades connecting varied glycemic indicators to coagulation remain uncharacterized. We hypothesize that fasting plasma glucose (FPG), fructosamine (FA), and glycated hemoglobin (HbA1c) distinctly influence D-dimer-defined hypercoagulability through specific upstream multisystem pathways, primarily mediated by systemic inflammation.

AIM

To delineate the mechanisms linking multi-time-scale glycemic indicators to hypercoagulability via multisystem pathways using structural equation modeling (SEM).

METHODS

In a retrospective cross-sectional study at Hangzhou Seventh People’s Hospital, we extracted electronic medical records of 4976 psychiatric inpatients. FPG, FA, and HbA1c were parallel exposures. We constructed a serial SEM to evaluate pathways connecting glycemia, multisystem nodes (lipids, blood pressure, inflammation, nutrition), coagulation substrates, and D-dimer hypercoagulability, adjusting for confounders.

RESULTS

Among 4976 psychiatric inpatients (D-dimer positivity: 14.4%), the SEM demonstrated excellent fit, explaining 18.4% of D-dimer variance. FPG exerted a positive total effect on D-dimer positivity [standardized beta coefficient (Std.β) = 0.112], whereas FA showed a robust negative effect (Std.β = -0.196); Effects related to HbA1c were weak and largely failed to reach statistical significance. Mechanistically, glycemic indicators drove hypercoagulability predominantly via multisystem pathways. The inflammatory pathway most strongly influenced the coagulation substrate node (Std.β = 0.42), which inversely correlated with D-dimer (Std.β = -0.09). These mechanistic patterns remained highly consistent across schizophrenia, bipolar disorder, and major depressive disorder subgroups, alongside multiple sensitivity analyses. Finally, multi-indicator joint modeling outperformed single-indicator models, confirming robust study conclusions.

CONCLUSION

Glycemic indicators in psychiatric inpatients influence hypercoagulability via multisystem pathways. FPG increases risk, FA decreases it; HbA1c effects are limited. Inflammation is central to clinical risk assessment.

Key Words: Major psychiatric disorders; Fasting plasma glucose; Fructosamine; Hemoglobin A1c; D-dimer; Structural equation modeling

Core Tip: This study innovatively applies structural equation modeling to reveal how glycemic indicators across different time scales influence hypercoagulability, marked by D-dimer, in psychiatric inpatients. Fasting plasma glucose actively promotes hypercoagulability via inflammatory pathways. Conversely, fructosamine robustly protects against it, acting as a surrogate for nutritional and anticoagulant reserves rather than mere glycemic control. Glycated hemoglobin showed negligible direct acute effects. These findings shift the paradigm of psychiatric coagulopathy from a simplistic metabolic complication to a multi-system cascade, highlighting the necessity of integrated metabolic, inflammatory, and nutritional clinical assessments.

INTRODUCTION

Major psychiatric disorders (MPDs), primarily encompassing schizophrenia (SCZ), bipolar disorder (BD), and major depressive disorder (MDD)[1], constitute a leading cause of the global disease burden, exerting profound and enduring impacts on patients’ daily lives, social functioning, and quality of life[2]. Venous thromboembolism (VTE) is recognized as one of the five major vascular diseases[3]. Studies have established a correlation between MPDs and VTE risk[4]; patients with psychiatric disorders may face a two-fold to three-fold higher risk of VTE compared to controls[5]. The pathogenesis of this hypercoagulable state stems from a complex interplay of multiple factors. Crucially, compelling evidence highlights the profound inflammatory underpinnings of MPDs. Extensive meta-analytical data demonstrate that acute-phase reactants, particularly C-reactive protein (CRP), and pro-inflammatory cytokines are significantly elevated in patients with SCZ, BD, and MDD. This reflects a systemic low-grade inflammatory state that is intrinsically linked to the underlying pathophysiology and severity of psychiatric symptoms[6,7]. Due to the retrospective, real-world nature of the study, systematic collection of detailed prospective data was not feasible. These psychiatric symptom-associated stress and neuroinflammatory responses, infections[3], and drug-induced metabolic disturbances (e.g., obesity, diabetes, and hyperlipidemia[8]) can perturb the dynamic equilibrium of the coagulation-fibrinolysis system via multiple pathways[9]. Compounding this intrinsic susceptibility, behavioral and iatrogenic factors - including a sedentary lifestyle, reduced mobility, dehydration, physical restraint, and antipsychotic-related adverse effects[10,11] - further elevate thrombotic risk, thereby synergistically driving the onset and progression of hypercoagulability. D-dimer, the most extensively utilized clinical biomarker of coagulation and fibrinolysis activation, is particularly critical in the long-term inpatient management of psychiatric disorders[12]. Elevated D-dimer levels are associated with prolonged hospitalization, exacerbated symptoms, poorer prognosis, and increased systemic inflammatory burden[4]. Consequently, within the psychiatric inpatient setting, thoroughly elucidating the drivers of abnormal D-dimer elevation holds significant clinical translational value for the early identification of VTE risk.

In recent years, dysregulated glycemic metabolic homeostasis has been recognized as a key upstream mechanism in VTE pathogenesis[13,14]. However, in complex psychiatric clinical scenarios, the relationship between glycemic variability across different time scales and coagulation function remains poorly elucidated. Fasting plasma glucose (FPG) reflects short-term, instantaneous glucose levels and is highly susceptible to stress, acute disease fluctuations, and insulin resistance induced by antipsychotics (particularly second-generation antipsychotics[15]). It rapidly affects coagulation pathways primarily by inducing inflammatory cytokine release, enhancing oxidative stress, and promoting platelet activation[16]. Fructosamine (FA) reflects the level of glycated serum proteins; with a half-life of 2-3 weeks, it represents a shorter period of glycemic control[17,18], effectively capturing glycemic fluctuations during psychiatric hospitalization. Albumin-corrected FA can further eliminate confounding from conditions such as hypoalbuminemia and malnutrition, thereby enhancing the accurate reflection of true glycemic levels[17]. FA may be associated with protein glycation, low-grade inflammatory responses, and vascular endothelial dysfunction[19]. Glycated hemoglobin (HbA1c) reflects long-term glycemic control over the preceding three months[20], serving as a vital marker of chronic metabolic burden in psychiatric patients. Long-term HbA1c elevation increases procoagulant risk by exacerbating lipid metabolism disorders, endothelial injury, and chronic metabolic cumulative effects[21]. Extensive research demonstrates that these multi-time-scale glycemic indicators exhibit upward trends across disorders such as SCZ, BD, and MDD[22]. This is closely associated with pharmacological treatments[23], lifestyle modifications[24], inflammation-stress responses[25], and the inherent metabolic susceptibility of the disorders themselves. However, previous findings present a complex and inconsistent pattern: While some studies suggest a positive correlation between elevated HbA1c or FPG and D-dimer[26,27], others report that FA does not consistently show a positive shift with coagulation in certain populations[28]. This suggests that the impact of blood sugar on the coagulation system may not be a simple linear relationship, but rather a complex network involving lipid metabolism, blood pressure, inflammatory cascades, nutritional status, and endothelial function.

Based on the above background, this study proposes the following hypothesis: In psychiatric inpatients, glycemic exposures across different temporal dimensions (FPG, FA, HbA1c) not only directly impact hypercoagulability but also indirectly modulate D-dimer levels via specific multisystem physiological pathways (encompassing blood lipids, blood pressure, inflammation, nutrition, and coagulation substrates). Given that conventional linear regression struggles to isolate complex variable interactions, this study innovatively employs structural equation modeling (SEM) to construct and validate a multipath network, aiming to quantify the direct effect, indirect effect, and total effects of different glycemic indicators on the high coagulation risk. This work will help to reveal the pathological mechanisms of the “three glycemic exposures - multisystem pathways - downstream coagulation nodes - D-dimer hypercoagulability” in patients with MPDs from a systems biology perspective, providing a theoretical basis for precise clinical interventions (Figure 1).

Figure 1 Multisystem pathway model linking three glycemic indicators to D-dimer hypercoagulability in psychiatric inpatients. Three mutually adjusted glycemic indicators (fasting plasma glucose, fructosamine, and glycated hemoglobin) were modeled as competing exposures influencing four upstream multi-system pathways composites (lipid, blood pressure, inflammation, and nutrition), which in turn acted through two proximal coagulation mediators-Coag (Substrate) and Coag (Time)-to affect hypercoagulable status (D-dimer > 0.5 μg/mL). BP: Blood pressure; FA: Fructosamine; FPG: Fasting plasma glucose; HbA1c: Glycated hemoglobin.

MATERIALS AND METHODS

Study design and participants

This study is based on the electronic medical record system. We centrally extracted data spanning from January 2018 to June 2025 through the Information Center of Hangzhou Seventh People’s Hospital. Leveraging the hospital’s electronic medical record system and clinical research data platform, we integrated structured data, including admission assessments, demographics, prior physical and psychiatric medical histories, psychiatric diagnoses, laboratory results, and vital signs. To construct comparable cross-sectional observations centered on the time of admission, we established the admission date as a uniform temporal anchor and utilized a data collection window of “2 days prior to admission to 2 days post-admission (totaling 5 days, including the day of admission)”. If multiple tests were recorded for the same indicator within this window, we retained only the earliest test result to minimize the potential impact of early inpatient therapeutic interventions on laboratory parameters. For patients who had multiple hospitalization records during the study period, to avoid duplicate inclusion and the statistical bias caused by repeated observations, the main analysis of this study only retained their first hospitalization record. This data curation strategy aimed to enhance the standardization and analytical consistency of real-world clinical data (Figure 2).

Figure 2 Flow diagram of patient selection. This study reviewed the electronic medical records of patients admitted to Hangzhou Seventh People’s Hospital between January 2018 and June 2025. Patients with complete data on all three glycemic indicators - fasting plasma glucose, fructosamine, and glycated hemoglobin - were initially identified. Records with diagnoses other than schizophrenia, bipolar disorder, or major depressive disorder were excluded. To ensure independent observations and data quality, duplicate hospitalization records, admissions with an excessively long length of stay (> 100 days), and records containing implausible or erroneous extreme values were further excluded. The final analytical cohort consisted of 4976 patients. FA: Fructosamine; FPG: Fasting plasma glucose; HbA1c: Glycated hemoglobin.

We included inpatients with psychiatric diagnoses of SCZ, BD, or MDD, established by psychiatrists according to International Statistical Classification of Diseases and Related Health Problems, 10^th Revision diagnostic criteria. All included subjects had to have completed tests for FPG, FA, and HbA1c within the predefined time window. To mitigate significant interference from potential confounders or specific physiological/pathological states on coagulation and inflammatory markers, we excluded patients with clearly documented acute major thrombotic events upon admission (e.g., pulmonary embolism, myocardial infarction, or acute cerebral infarction), active malignancies (especially hematological malignancies), pregnancy or perinatal status, and severe hepatic or renal failure.

SEM analysis design

We employed a serial SEM approach to test the potential mechanistic chain of “three glycemic indicators - upstream multisystem pathways - downstream coagulation nodes- D-dimer hypercoagulable state”. We specified the primary model as a partial mediation model, analyzing the path effects of glycemic indicators on coagulation status via multisystem pathways under conditionally controlled covariates (Table 1).

Table 1 Composite indices, observed variables, and theoretical directions in the main structural equation model¹.

Composite index	Observed variable	Theoretical direction
Lipid pathway (lipid index)	Triglycerides (TG)	Positive: Higher TG → higher lipid index → theoretically higher hypercoagulability/thrombotic risk
	Total cholesterol (TC)	Positive: Higher TC → higher lipid index → theoretically higher hypercoagulability/thrombotic risk
	Low-density lipoprotein cholesterol (LDL-C)	Positive: Higher LDL-C → higher lipid index → theoretically higher hypercoagulability/thrombotic risk
	High-density lipoprotein cholesterol (HDL-C)	Negative: Higher HDL-C → lower lipid index (protective) → theoretically lower hypercoagulability risk
Blood pressure pathway (BP index)	Systolic blood pressure (SBP)	Positive: Higher SBP → higher BP index → theoretically higher hypercoagulability risk
	Diastolic blood pressure (DBP)	Positive: Higher DBP → higher BP index → theoretically higher hypercoagulability risk
Inflammation pathway (inflamm index)	C-reactive protein (CRP)	Positive: Higher CRP → higher inflamm index→ theoretically higher hypercoagulability risk
	Serum amyloid A (SAA)	Positive: Higher SAA → higher inflamm index → theoretically higher hypercoagulability risk
	Total white blood cell count (WBC)	Positive: Higher WBC → higher inflamm index → theoretically higher hypercoagulability risk
Nutritional pathway (nutrition index)	Hemoglobin (HB, marker of anemia and nutritional status)	Negative: Higher HB → lower nutrition index (better nutrition) → theoretically lower hypercoagulability risk
	Prealbumin (PAB, short-term nutritional marker)	Negative: Higher PAB → lower nutrition index (better nutrition) → theoretically lower hypercoagulability risk
	Total protein (TP, hepatic synthetic function and nutritional status)	Negative: Higher TP → lower nutrition index (better synthetic/nutritional status) → theoretically lower hypercoagulability risk
Coagulation substrate [coag (substrate) index]	Fibrinogen (FIB)	Positive: Higher FIB → higher endocoag index → theoretically higher hypercoagulability risk
	Platelet count (PLT)	Positive: Higher PLT → higher endocoag index → theoretically higher hypercoagulability risk
Coagulation function/time [coag (time) index]	Activated partial thromboplastin time (APTT)	Negative: Longer APTT → stronger anticoagulation/insufficient coagulation factors → theoretically lower hypercoagulability risk
	Prothrombin time (PT)	Negative: Longer PT → stronger anticoagulation/insufficient coagulation factors → theoretically lower hypercoagulability risk
	Thrombin time (TT)	Negative: Longer TT → more pronounced anticoagulation/fibrinolysis → theoretically lower hypercoagulability risk

¹Composite indices were defined to represent major biological pathway domains hypothesized to link glycemic dysregulation with hypercoagulability in the main structural equation model. Theoretical directions indicate the expected contribution of each observed variable to its corresponding composite index based on established physiological and clinical evidence. For most indices, higher values were specified to reflect greater pathway activity and, consequently, a higher theoretical propensity toward hypercoagulability. The nutrition index was oriented inversely, such that higher values indicate poorer nutritional status and potentially increased prothrombotic risk. Prior to structural equation model estimation, all observed variables were standardized and directionally aligned according to these theoretical assumptions to ensure interpretability and consistency of the composite indices.

TG: Triglycerides; TC: Total cholesterol; LDL-C: Low-density lipoprotein cholesterol; HDL-C: High-density lipoprotein cholesterol; BP: Blood pressure; SBP: Systolic blood pressure; DBP: Diastolic blood pressure; CRP: C-reactive protein; SAA: Serum amyloid A; WBC: Total white blood cell count; HB: Hemoglobin; PAB: Prealbumin; TP: Total protein; FIB: Fibrinogen; PLT: Platelet count; APTT: Activated partial thromboplastin time; PT: Prothrombin time; TT: Thrombin time.

Exposure layer: FPG, FA, and HbA1c were incorporated into the SEM as primary exposure variables. In the SEM analysis, these three indicators entered the model simultaneously as mutually adjusted exposure variables, adjusting for each other to estimate their conditional associations. Given that this study focused on the independent contributions of each indicator to the mechanistic chain rather than the correlation structure among the three sugars, we did not parameterize the exogenous covariance among them in the structural equation.

Upstream pathway layer: Multisystem mechanistic pathways were delineated using composite indices constructed from homologous physiological observation metrics to reflect systemic physiological states, rather than utilizing reflective latent variable measurement models. We first Z-score standardized, linearly synthesized, and re-standardized constituent metrics to obtain the final indices. The four upstream pathways included the lipid index [triglycerides, total cholesterol, and low-density lipoprotein cholesterol weighted positively; high-density lipoprotein (HDL) weighted negatively], blood pressure index (systolic blood pressure and diastolic blood pressure), inflammation index (CRP, serum amyloid A, and white blood cell), and nutrition index (hemoglobin, prealbumin, and total protein weighted negatively, such that a higher index indicated poorer nutritional status). In the SEM, these four pathway indices functioned as parallel mediation modules regressed on the three glycemic indicators, allowing for residual correlations among pathways to account for potential unmeasured common background factors.

Downstream coagulation layer: Two proximal mechanistic nodes were constructed as downstream mediators. These included the coagulation substrate index (derived from fibrinogen and platelet count) and the coagulation time/function index (synthesized via reverse weighting of activated partial thromboplastin time, prothrombin time, and thrombin time). Considering that these two indices reflected different functional dimensions of the coagulation system, we pre-specified their residual correlation in the model to account for unmeasured common physiological foundations. We did not specify direct paths from glycemic indicators to downstream coagulation nodes in the primary model, emphasizing the hypothesis that glycemic effects on the proximal coagulation system were primarily mediated indirectly via multisystem pathways.

Outcome layer: The primary outcome was the hypercoagulable state defined by D-dimer (D-dimer > 0.5 μg/mL). We treated this outcome as a binary categorical variable in the SEM, utilizing weighted least squares mean and variance adjusted estimation. We did not impute missing values, adhering to lavaan’s default missing data handling strategy under weighted least squares mean and variance estimation (non-full information maximum likelihood), and utilized a probit link function in the model. We retained direct paths from the three glycemic indicators and upstream multisystem pathways to the hypercoagulable state to prevent over-constraining the model and improve fit feasibility.

Covariates and confounding control: Covariates included age, sex, body mass index, smoking history, alcohol consumption history, prior medical history of hypertension, diabetes, and hyperlipidemia, duration of psychiatric illness (years), and abnormal body temperature status (defined as temperature ≥ 37.4 °C at any point within the predefined time window). We incorporated covariates into the SEM as exogenous variables, employing a conditional estimation strategy for confounding control (implemented in lavaan by setting fixed.x = TRUE). Specifically, we regressed covariates solely onto the upstream multisystem pathway indices, downstream coagulation nodes, and outcome variables as predictors, without modeling the covariance structure among the covariates themselves, thereby estimating path effects conditionally upon the specified covariates (Supplementary material).

Sensitivity analyses

We employed a one-at-a-time adjustment strategy within the same SEM framework to conduct sensitivity analyses. By altering only one analytical dimension per iteration while keeping other model specifications consistent with the primary model, we avoided excessive expansion of the model space and enhanced the interpretability of the results. First, at the structural level, we assessed model robustness by adjusting constraint intensity, testing both a relaxed partial mediation model (allowing direct paths from glycemic indicators to coagulation nodes) and a full mediation model (constraining direct paths from glycemic indicators and upstream pathways to the outcome to zero). Second, regarding the outcome definition, we replaced the binary D-dimer hypercoagulable state with a continuous log-transformed D-dimer metric. This was assessed by means of the maximum likelihood with robust standard errors estimation, with full information maximum likelihood being used to deal with missing data within this framework. Third, at the indicator operationalization level, we replaced or extended the operationalization of the construction methods for various system pathway indices. The lipid pathway used the atherogenic index of plasma instead of the traditional lipid composite index. For the inflammation pathway, we evaluated a synthesized inflammation index (after symmetric truncation of inflammatory markers), an expanded inflammatory marker set (including neutrophil count, monocyte count, and lactate dehydrogenase in addition to the basic markers), and the systemic inflammatory response index. The controlling nutritional status score and the prognostic nutritional index were used as surrogate indicators for the nutrition pathway. Fourth, concerning the confounding control, we added extensive indices on liver and kidney function to assess how the model path estimations would be affected if there were organ function confounding with the exogenous variables fixed. Fifth, at the exposure configuration level, we assessed the stability of key path estimates obtained by using various exposure representations: Single-glycemic-indicator models, an albumin-corrected FA substitution model, and a multi-indicator synthetic exposure model (Table 2 and Supplementary material).

Table 2 Additional composite indices and observed variables included in sensitivity analyses¹.

Composite index	Observed variable	Theoretical direction
Lipid pathway (lipid index)	Atherogenic index of plasma, AIP = log10(TG/HDL-C)	Positive: Higher AIP → greater atherogenic burden → theoretically higher hypercoagulability risk
Inflammation pathway (inflamm index)	Absolute neutrophil count (NEU)	Positive: Higher NEU → higher inflammatory/stress response → theoretically higher hypercoagulability risk
	Absolute monocyte count (MONO)	Positive: Higher MONO → higher chronic inflammation/monocyte–macrophage activation → theoretically higher hypercoagulability risk
	Lactate dehydrogenase (LDH)	Positive: Higher LDH → more tissue damage/inflammatory burden → theoretically higher hypercoagulability risk
	Systemic inflammation response index (SIRI = NEU × MONO / LYM)	Positive: Higher SIRI → higher immune-inflammatory burden → theoretically higher hypercoagulability risk
Nutritional/ immune sensitivity indices	CONUT score (albumin + total cholesterol + lymphocyte count)	Positive: Higher CONUT score → worse malnutrition → theoretically worse outcomes/higher hypercoagulability risk
	Prognostic nutritional index (PNI = ALB + 5 × LYM)	Negative: Higher PNI (better albumin and lymphocyte status) → better nutritional/immune status → theoretically lower hypercoagulability risk
Composite covariates	Observed variable	Theoretical direction
Hepatic burden domain (hepatic index)	Alanine aminotransferase (ALT)	Positive: Higher ALT → greater hepatocellular injury → theoretically higher hypercoagulability risk
	Aspartate aminotransferase (AST)	Positive: Higher AST → greater tissue injury → theoretically higher hypercoagulability risk
	Gamma-glutamyl transferase (GGT)	Positive: Higher GGT → greater metabolic/cholestatic burden → theoretically higher hypercoagulability risk
	Alkaline phosphatase (ALP)	Positive: Higher ALP → greater biliary/bone metabolic burden → theoretically higher hypercoagulability risk
	Total bilirubin (TBIL)	Positive: Higher TBIL → greater hepatic/cholestatic dysfunction → theoretically higher hypercoagulability risk
	Direct bilirubin (DBIL)	Positive: Higher DBIL → greater hepatobiliary dysfunction → theoretically higher hypercoagulability risk
Renal burden domain (renal index)	Blood urea nitrogen (BUN)	Positive: Higher BUN → worse renal function/protein metabolism → theoretically higher hypercoagulability risk
	Serum creatinine (Cr)	Positive: Higher Cr → worse renal function → theoretically higher hypercoagulability risk

¹The composite indices and observed variables listed in this table were incorporated exclusively in sensitivity analyses to evaluate the robustness of the main structural equation model. These indices were specified a priori as alternative or complementary operationalizations of key biological domains, including lipid-related atherogenic burden, inflammatory and immune activation, nutritional and immunological status, and hepatic and renal function. Theoretical directions indicate the expected contribution of each observed variable to its corresponding index based on established pathophysiological and clinical evidence. In general, higher index values were defined to represent greater pathway burden and a theoretically higher propensity toward hypercoagulability. The prognostic nutritional index was oriented inversely, such that higher values indicate better nutritional and immune status. Hepatic and renal indices were treated as additional composite covariates in sensitivity analyses and were not modeled as mediators in the primary structural equation model framework.

AIP: Atherogenic index of plasma; TG: Triglycerides; HDL-C: High-density lipoprotein cholesterol; NEU: Absolute neutrophil count; MONO: Absolute monocyte count; LDH: Lactate dehydrogenase; SIRI: Systemic inflammation response index; LYM: Lymphocyte count; CONUT: Controlling nutritional status; PNI: Prognostic nutritional index; ALB: Albumin; ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; GGT: Gamma-glutamyl transferase; ALP: Alkaline phosphatase; TBIL: Total bilirubin; DBIL: Direct bilirubin; BUN: Blood urea nitrogen; Cr: Creatinine.

Statistical analysis and visualization

We executed all statistical analyses in the R statistical environment (R Foundation for Statistical Computing, Vienna, Austria; version 4.5.1), primarily via the RStudio platform. Data preprocessing was conducted using various R packages, including readxl, dplyr, tidyr, and stringr. This involved stripping out duplicate records, translating non-numeric test results with symbols such as “ > “ to numeric values with predefined rules, fixing obvious data entry mistakes, and uniformising the encoding of missing test values. We generated data visualizations predominantly using ggplot2. We first conducted descriptive statistical analyses on the overall sample and individual diagnostic subgroups (SCZ, BD, and MDD). We expressed continuous variables as mean ± SD or median (interquartile range), depending on their distribution characteristics, and presented categorical variables as frequencies (percentages). We performed pairwise correlation analyses of laboratory parameters, covariates, and multisystem pathway composite indices with a Spearman rank correlation test and plotted a correlation matrix heatmap to depict the overall correlation structure. A Mantel was used to assess the global correlations between glycemic indicators and pathway indicator matrices. The SEM was conducted using the lavaan package, and standardized path coefficients were reported, as well as direct, indirect, and total effects. We have also presented the explained variance (R²) for each endogenous variable. The outer bootstrap resampling method (500 iterations) was used to estimate standardized path coefficients [standardized beta coefficient (Std.β)] for the main model and their 95% confidence intervals, with the percentile method being used to construct confidence intervals. Forest plots were used to visualize important structural pathways. Model goodness-of-fit was thoroughly examined with multiple indices such as the comparative fit index (CFI), Tucker-Lewis index, Root Mean Square Error of Approximation (RMSEA), Standardized Root Mean Square Residual, and degrees of freedom (df). To comprehensively evaluate the robustness of the sensitivity analysis results, we used path direction consistency as the core criterion, and in combination, the magnitude of the change of the standardized path coefficients (|ΔStd.β|). Specifically, we considered a change of |ΔStd.β| < 0.20 as an acceptable variation range, serving as an auxiliary metric to judge the stability of key paths.

RESULTS

Descriptive statistics

We included a total of 4976 psychiatric inpatients in this study. Data for the three glycemic indicators were complete; the mean FPG was 6.5 ± 3.0 mmol/L, FA was 1.7 ± 0.4 mmol/L, and HbA1c was 5.7 ± 1.0%. The missing rate for D-dimer was 13.3%, and among those with available data, hypercoagulability (D-dimer > 0.5 μg/mL) accounted for 14.4%. Laboratory indicators used to construct the multisystem pathway indices had generally low missing rates, though the missing rate for coagulation-related tests (fibrinogen, activated partial thromboplastin time, prothrombin time, thrombin time) was approximately 12.9%, which was highly consistent with the D-dimer testing coverage. Among inflammatory markers, CRP had a relatively higher missing rate (12.3%), whereas white blood cell and serum amyloid A data were more complete. The mean age of the subjects was 41 ± 18 years, with males comprising 35.4%, and the mean body mass index was 23.1 ± 4.3 kg/m². The prevalences of prior hypertension, diabetes, and hyperlipidemia were 17.9%, 12.0%, and 4.2%, respectively, and approximately 18.3% of patients exhibited abnormal body temperature during early admission (Table 3).

Table 3 Descriptive summary of clinical variables and observed laboratory measures¹.

Variable	Unit	Missing (n)	Missing (%)	mean ± SD	Distribution
Glycemic exposures
FPG	mmol/L	0	0	6.5 ± 3.0	5.6 (5.0-6.8)
FA	mmol/L	0	0	1.7 ± 0.4	1.6 (1.5-1.8)
HbA1c	%	0	0	5.7 ± 1.0	5.4 (5.2-5.8)
FA-ALBcorr		4	0.08	3.90 ± 0.82	3.72 (3.42-4.12)
Outcomes
D-dimer	μg/mL	663	13.32	0.3 ± 0.8	0.1 (0.1-0.3)
log(D-dimer)		663	13.32	0.24 ± 0.29	0.14 (0.06-0.29)
D-dimer (+)		663	13.32		622 (14.4)
Components/derived measures
TG	mmol/L	3	0.06	1.7 ± 1.4	1.3 (0.9-2.0)
TC	mmol/L	70	1.41	4.5 ± 1.0	4.4 (3.8-5.0)
LDL-C	mmol/L	3	0.06	2.6 ± 0.9	2.5 (2.0-3.1)
HDL-C	mmol/L	3	0.06	1.3 ± 0.4	1.3 (1.1-1.5)
SBP	mmHg	1	0.02	125 ± 17	124 (114-135)
DBP	mmHg	1	0.02	80 ± 11	80 (72-87)
Hb	g/L	33	0.66	135 ± 16	134 (125-146)
PAB	mg/L	5	0.1	286 ± 73	278 (236-327)
TP	g/L	4	0.08	70 ± 6	71 (66-74)
ALB	g/L	4	0.08	44 ± 4	44 (41-47)
FIB	g/L	641	12.88	2.7 ± 0.7	2.6 (2.2-3.0)
PLT	× 109/L	33	0.66	239 ± 69	233 (193-281)
APTT	seconds	641	12.88	28.8 ± 4.9	28.1 (25.8-31.0)
PT	seconds	641	12.88	11.4 ± 1.3	11.3 (10.8-11.9)
TT	seconds	641	12.88	15.3 ± 3.9	16.8 (11.1-18.6)
ALT	U/L	0	0	25 ± 26	17 (12-27)
AST	U/L	0	0	24 ± 19	19 (16-25)
GGT	U/L	73	1.47	25 ± 30	17 (12-27)
ALP	U/L	5	0.1	81 ± 32	75 (61-93)
TBIL	µmol/L	1	0.02	10.5 ± 6.3	8.9 (6.4-12.6)
DBIL	µmol/L	2	0.04	3.1 ± 2.3	2.5 (1.8-3.6)
BUN	mmol/L	1	0.02	4.8 ± 1.8	4.5 (3.7-5.6)
Cr	µmol/L	1	0.02	63.6 ± 17.1	61.0 (53.0-72.0)
WBC	× 109/L	33	0.66	7.3 ± 2.4	6.9 (5.6-8.6)
CRP	mg/L	612	12.3	3.1 ± 9.0	0.8 (0.5-2.5)
SAA	mg/L	37	0.74	13.6 ± 38.1	5.8 (3.9-9.5)
LDH	U/L	3	0.06	173 ± 47	164 (144-190)
NEU	× 109/L	101	2.03	4.8 ± 2.1	4.3 (3.3-5.8)
MONO	× 109/L	101	2.03	0.5 ± 0.2	0.5 (0.4-0.6)
LYM	× 109/L	101	2.03	1.9 ± 0.7	1.8 (1.4-2.3)
AIP		3	0.06	0.02 ± 0.31	0.00 (-0.20-0.23)
SIRI		101	2.03	1.56 ± 1.57	1.12 (0.71-1.84)
CONUT		105	2.11
Normal (0-1)					4144 (85.1)
Mild malnutrition (2-4)					721 (14.8)
Moderate severe (5-8)					6 (0.1)
PNI		105	2.11	53.4 ± 5.5	53.2 (49.6-56.8)
Covariates
Age	years	0	0	41 ± 18	39 (26-56)
Gender, male		0	0		1762 (35.4)
BMI	kg/m2	0	0	23.1 ± 4.3	22.7 (20.1-25.5)
Hypertension (+)		0	0		893 (17.9)
Diabetes (+)		0	0		595 (12.0)
Hyperlipidemia (+)		0	0		210 (4.2)
Smoking Status		0	0
No history					4189 (84.2)
Current					705 (14.2)
Former					82 (1.6)
Alcohol Status		0	0
No history					4312 (86.7)
Current					629 (12.6)
Former					35 (0.7)
Body temperature (+)		0	0		912 (18.3)
Psychiatric duration	years	0	0	9 ± 10	5 (1-12)

¹Continuous variables are presented as mean ± SD and median (interquartile range), whereas categorical variables are reported as number (percentage). Percentages were calculated based on non-missing observations for each variable. D-dimer positivity [D-dimer (+)] was defined as a plasma D-dimer concentration > 0.5 μg/mL. Log-transformed D-dimer values were calculated as log(D-dimer + 1) to accommodate zero values and reduce skewness. Albumin-corrected fructosamine (FA-ALBcorr) denotes albumin-corrected fructosamine. The atherogenic index of plasma was calculated as log₁₀(triglyceride to high-density lipoprotein cholesterol ratio), and the systemic inflammation response index was defined as neutrophil × monocyte/ lymphocyte counts. Nutritional status was assessed using the Controlling Nutritional Status score and categorized as normal (0-1), mild malnutrition (2-4), or moderate to severe malnutrition (5-8). Abnormal body temperature [body temperature (+)] was defined as any recorded body temperature > 37.3 °C within the first 3 days after hospital admission and was treated as a binary variable (1 = criterion met; 0 = not met).

FPG: Fasting plasma glucose; FA: Fructosamine; HbA1c: Glycated hemoglobin; FA-ALBcorr: Albumin-corrected fructosamine; D-dimer (+): D-dimer positivity; TG: Triglycerides; TC: Total cholesterol; LDL-C: Low-density lipoprotein cholesterol; HDL-C: High-density lipoprotein cholesterol; SBP: Systolic blood pressure; DBP: Diastolic blood pressure; Hb: Hemoglobin; PAB: Prealbumin; TP: Total protein; ALB: Albumin; FIB: Fibrinogen; PLT: Platelet count; APTT: Activated partial thromboplastin time; PT: Prothrombin time; TT: Thrombin time; ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; GGT: Gamma-glutamyl transferase; ALP: Alkaline phosphatase; TBIL: Total bilirubin; DBIL: Direct bilirubin; BUN: Blood urea nitrogen; Cr: Creatinine; WBC: White blood cell count; CRP: C-reactive protein; SAA: Serum amyloid A; LDH: Lactate dehydrogenase; NEU: Absolute neutrophil count; MONO: Absolute monocyte count; LYM: Lymphocyte count; AIP: Atherogenic index of plasma; SIRI: Systemic inflammation response index; CONUT: Controlling nutritional status; PNI: Prognostic nutritional index; BMI: Body mass index.

Upon stratifying by diagnosis, we categorized the cohort into 1935 patients with SCZ, 1111 with BD, and 1,930 with MDD. The overall levels of the three glycemic indicators were similar across groups, although the SCZ group had a slightly higher FPG (7.0 ± 3.5 mmol/L). The proportions of D-dimer positivity were 15.1% in the SCZ group, 11.2% in the BD group, and 15.7% in the MDD group, demonstrating no substantial overall difference. Demographically, BD patients were younger (33 ± 16 years), MDD patients were relatively older (46 ± 20 years), and the SCZ group had a higher proportion of males (41.8%). Psychiatric disease duration was relatively longer in the SCZ and BD groups (10 ± 11 years and 9 ± 10 years, respectively). Regarding inflammatory and coagulation-related indicators, the SCZ group exhibited generally higher inflammation levels, whereas the MDD group had relatively lower nutritional indices and the highest prevalence of hypertension (26.2%). Overall, while certain metabolic and inflammatory differences existed across diagnostic subgroups, the primary variable distributions largely overlapped, supporting subsequent stratified and overall joint analyses (Supplementary material 1).

The three glycemic indicators exhibited specific correlation structures with multisystem biological characteristics, though the overall correlation strengths were moderate and presented differential patterns. FPG demonstrated more prominent correlations with blood pressure and inflammation-related indicators, whereas FA showed more consistent overall correlations with metabolic and selected clinical characteristics. In contrast, the correlation strengths for HbA1c with most indicators were relatively weak. Mantel test results further indicated a global correlation between the glycemic indicators and the multisystem pathway characteristic matrices, yet the correlation patterns varied across different glycemic phenotypes. Overall, the correlation structure supported the presence of systemic links between glycemic status and the inflammation-coagulation network, which were not predominantly driven by a single indicator (Figure 3, Supplementary materials 2-4).

Figure 3 Associations between glycemic markers and multisystem pathway features in the overall cohort and diagnostic subgroups of psychiatric inpatients. A: Laboratory biomarkers; B: Covariate features; C: Pathway-level indices. A-C summarize the relationships between three glycemic indicators - fasting plasma glucose, fructosamine, and glycated hemoglobin - and multilevel biological features spanning metabolic, cardiovascular, inflammatory, nutritional, and coagulation domains. In each panel, lower-triangle heatmaps display pairwise Spearman correlation coefficients among observed variables or composite pathway indices (red indicates positive correlations and blue indicates negative correlations, with numeric coefficients shown in each cell). Curved links on the right represent associations between each glycemic marker and the corresponding feature matrix assessed using Mantel tests (maximum matrix size constrained to n = 300 and significance evaluated using 999 permutations). Line thickness reflects the absolute Mantel correlation coefficient (|r|), and link color denotes statistical significance after Benjamini-Hochberg false discovery rate correction: Green, q < 0.01; yellow, q = 0.01-0.05; grey, q ≥ 0.05. A focuses on laboratory biomarkers constituting the composite pathway indices or their sensitivity-analysis components, including markers related to lipid metabolism, blood pressure, inflammation, nutritional status, and coagulation function. B presents correlations with covariate features incorporated in the structural models, including demographic characteristics, lifestyle factors, clinical comorbidities, psychiatric illness duration, body temperature status, and hepatic and renal function indicators. C visualizes pathway-level composite indices aligned with the structural modeling framework, including lipid, blood pressure, inflammatory, nutritional, hepatic, renal, and coagulation pathway constructs, together with downstream coagulation dimensions - coagulation substrate and coagulation time/function - as well as D-dimer outcomes. Visually, the heatmaps show clustered correlation patterns within biologically related domains, with relatively stronger intra-domain correlations among laboratory biomarkers (A) and more heterogeneous correlation structures among covariate features (B) and pathway-level composite indices (C).

Main model analysis

In the overall sample, the SEM demonstrated excellent goodness-of-fit (CFI = 0.990, Tucker-Lewis index = 0.966, RMSEA = 0.041, Standardized Root Mean Square Residual = 0.001), and the model converged successfully. In the diagnostic subgroup analyses, the fit indices across models were generally consistent with the main model; specifically, the CFIs for the SCZ, BD, and MDD subgroup models were 0.988, 0.983, and 0.989, respectively, and the RMSEAs were 0.037, 0.049, and 0.053, indicating that the model structure possessed good fit consistency across different diagnostic populations. The coefficients of determination (R²) revealed that, in the overall sample, the nutritional pathway had the highest explanatory power among upstream pathways (R² = 0.291), followed by the lipid (0.207) and blood pressure pathways (0.186), while the inflammatory pathway was relatively lower (0.084). Among downstream nodes, the R² for the coagulation substrate index was 0.265, whereas it was lower for the coagulation time index (0.048). The overall explanatory power for D-dimer positivity was 0.184. In the diagnostic subgroup analyses, the explanatory power of each pathway was broadly consistent with the overall sample. The nutritional pathway maintained a high R² across all subgroups (R² = 0.265-0.295), the coagulation substrate index reached its highest R² in the MDD subgroup (R² = 0.303), and the coagulation time index was lowest in the SCZ subgroup (R² = 0.019). The R² for D-dimer positivity was slightly higher in the MDD subgroup (R² = 0.223) (Table 4).

Table 4 Explained variance (R²) of the composite index and outcome across diagnostic groups.

Composite index/outcome	Overall (ALL)	Schizophrenia (SCZ)	Bipolar disorder (BD)	Depression (MDD)
Lipid index	0.207	0.231	0.251	0.171
BP index	0.186	0.209	0.275	0.134
Inflammation index	0.084	0.064	0.094	0.096
Nutrition index	0.291	0.286	0.295	0.265
Coagulation substrate index	0.265	0.234	0.27	0.303
Coagulation time index	0.048	0.019	0.101	0.113
D-dimer (+)	0.184	0.191	0.185	0.223

ALL: Overall; SCZ: Schizophrenia; BD: Bipolar disorder; MDD: Major depressive disorder.

Mediation analysis revealed differentiated association patterns between the three glycemic indicators and D-dimer positivity. In the overall sample, FPG exerted a positive direct effect on D-dimer positivity (Std.β = 0.115), with a negligible negative indirect effect (Std.β = -0.004), resulting in a mild positive total effect (Std.β = 0.112). Conversely, FA demonstrated a stable negative association, characterized by a negative direct effect (Std.β = -0.192) accompanied by a minor negative indirect effect (Std.β = -0.004), yielding an overall negative total effect (Std.β = -0.196). In comparison, both the direct and indirect effects of HbA1c were weak, with confidence intervals frequently crossing zero, and the total effect lacked a clear statistical association (Std.β = 0.033). In the diagnostic subgroup analyses, the effect directions for FPG and FA were largely consistent with the overall sample, whereas HbA1c-related effects were weaker and less stable (Table 5).

Table 5 Standardized direct, indirect, and total effects of glycemic markers on D-dimer (+) across diagnostic groups (structural equation modeling).

Effect	Marker	Overall β (95%CI)	Std.β	SCZ β (95%CI)	SCZ Std.β	BD β (95%CI)	BD Std.β	MDD β (95%CI)	MDD Std.β
Direct	FPG	0.044 (0.017 to 0.071)	0.115	0.036 (-0.004 to 0.076)	0.11	0.057 (-0.007 to 0.121)	0.128	0.045 (-0.005 to 0.095)	0.094
Direct	FA	-0.640 (-0.879 to -0.402)	-0.192	-0.682 (-1.035 to -0.329)	-0.235	-0.665 (-1.257 to -0.073)	-0.172	-0.634 (-1.063 to -0.205)	-0.162
Direct	HbA1c	0.037 (-0.059 to 0.134)	0.035	0.026 (-0.126 to 0.179)	0.03	0.069 (-0.160 to 0.298)	0.059	0.085 (-0.090 to 0.259)	0.059
Indirect	FPG	-0.001 (-0.002 to -0.000)	-0.004	-0.001 (-0.002 to 0.000)	-0.003	0.000 (-0.003 to 0.003)	0	-0.001 (-0.003 to 0.002)	-0.001
Indirect	FA	-0.012 (-0.021 to -0.004)	-0.004	-0.003 (-0.010 to 0.004)	-0.001	-0.020 (-0.047 to 0.008)	-0.005	-0.019 (-0.043 to 0.005)	-0.005
Indirect	HbA1c	-0.002 (-0.004 to 0.001)	-0.002	-0.000 (-0.003 to 0.002)	0	-0.008 (-0.019 to 0.003)	-0.007	-0.016 (-0.028 to -0.003)	-0.011
Total	FPG	0.043 (0.016-0.070)	0.112	0.035 (-0.005 to 0.075)	0.107	0.056 (-0.008 to 0.121)	0.127	0.045 (-0.005 to 0.095)	0.093
Total	FA	-0.652 (-0.891 to -0.414)	-0.196	-0.685 (-1.037 to -0.332)	-0.236	-0.684 (-1.278 to -0.091)	-0.178	-0.653 (-1.081 to -0.225)	-0.167
Total	HbA1c	0.035 (-0.061 to 0.133)	0.033	0.026 (-0.126 to 0.178)	0.029	0.061 (-0.169 to 0.291)	0.052	0.069 (-0.104 to 0.242)	0.048

CI: Confidence interval; SCZ: Schizophrenia; BD: Bipolar disorder; MDD: Major depressive disorder; FPG: Fasting plasma glucose; FA: Fructosamine; HbA1c: Glycated hemoglobin.

Path analysis further indicated that glycemic indicators were associated with hypercoagulability predominantly via multisystem pathways. FPG correlated positively with the blood pressure and inflammation pathway indices (Std.β = 0.11 and 0.08, respectively). FA exhibited a stronger positive association with the lipid pathway (Std.β = 0.23) and a stable negative association with the nutritional pathway (Std.β = -0.32), whereas HbA1c correlated with all pathways were generally weak. At the downstream level, multisystem pathways exerted consistent effects on the coagulation substrate node, with the inflammatory pathway displaying the strongest effect (Std.β = 0.42); lipid and blood pressure pathways also showed positive associations (Std.β = 0.12 and 0.08, respectively). The coagulation substrate node was negatively associated with D-dimer positivity (Std.β = -0.09). This inverse association, while seemingly counterintuitive, likely captures a “high-turnover” consumptive coagulopathy state characterized by the rapid depletion of coagulation substrates, a phenomenon that will be explored in detail in the discussion section. By contrast, the effects of all pathways on the coagulation time node were generally weak. In the diagnostic subgroup analyses, path directions were largely consistent with the overall sample, though confidence intervals were wider. The effect of the inflammatory pathway on the coagulation substrate node remained robust across the SCZ, BD, and MDD subgroups (Std.β = 0.40, 0.41, and 0.45, respectively). Concurrently, the positive association of FPG and the negative association of FA with D-dimer were generally consistent across different diagnostic subgroups, while paths related to HbA1c exhibited weaker stability (Figures 4, 5, 6, and 7; Table 6; Supplementary materials 5-12).

Figure 4 Overview of the main serial structural equation model pathways and standardized bootstrap-estimated effects on D-dimer outcomes. A: Main serial structural equation modeling pathways; B: Bootstrap-estimated effects. The left (A) depicts the estimated serial structural equation modeling pathways linking three glycemic indicators - fasting plasma glucose, fructosamine, and glycated hemoglobin - to D-dimer outcomes in the main model. Glycemic indicators are modeled to relate to D-dimer through four upstream pathway-level indices (lipid index, blood pressure index, inflammation index, and nutrition index) and two downstream coagulation-related nodes capturing complementary aspects of hemostasis (coagulation substrate index and coagulation time index). Red arrows indicate positive standardized associations and blue arrows indicate negative associations. Solid lines denote statistically significant paths, whereas dashed lines indicate non-significant paths. Line width reflects the magnitude of the standardized coefficient (|Std.β|). Collectively, the structural component explains a non-trivial proportion of variance (R²) in key endogenous nodes: Lipid index = 0.207, blood pressure index = 0.186, inflammation index = 0.084, nutrition index = 0.291, coagulation substrate index = 0.265, coagulation time/function index = 0.048, and D-dimer = 0.184. The right (B) summarizes the corresponding standardized path estimates using forest plots with bootstrap-derived 95% confidence intervals [Std.β (95% confidence interval), 500 bootstrap resamples]. Effects are organized into three sequential layers aligned with the hypothesized serial pathway structure: (1) Glycemic indicators → upstream pathway indices; (2) Upstream pathway indices → downstream coagulation nodes; and (3) Direct effects on D-dimer. Points represent standardized coefficients, horizontal bars indicate bootstrap-derived 95% confidence intervals, and the vertical reference line denotes the null effect (Std.β = 0). Together, these combined visualization provides an integrated overview of the modeled multisystem associations linking glycemic markers, intermediate physiological pathways, and D-dimer outcomes within the main structural equation modeling framework. FA: Fructosamine; FPG: Fasting plasma glucose; HbA1c: Glycated hemoglobin; BP: Blood pressure; CI: Confidence interval.

Figure 5 Pathways linking fasting plasma glucose to D-dimer across four upstream pathways. A: Lipid pathway; B: Blood pressure pathway; C: Inflammation pathway; D: Nutrition pathway. A-D present system-specific path diagrams illustrating how fasting plasma glucose relates to D-dimer through upstream physiological pathway indices and a downstream coagulation substrate node. Each panel displays the full set of modeled pathways within the corresponding domain, including the direct association fasting plasma glucose → D-dimer and the indirect pathway fasting plasma glucose → upstream pathway index → coagulation (substrate) → D-dimer. Direct paths from each upstream pathway index to D-dimer are also shown, where specified. Panels correspond to four upstream pathway domains: (A) Lipid metabolism; (B) Blood pressure; (C) Inflammation; and (D) Nutrition. Arrow color denotes the direction of the standardized association (red = positive; cyan = negative), arrow thickness reflects the magnitude of the standardized coefficient (|Std.β|; 0-0.1, 0.1-0.3, > 0.3), and line type indicates statistical significance (solid = significant; dashed = not significant). FPG: Fasting plasma glucose; BP: Blood pressure.

Figure 6 Pathways linking fructosamine to D-dimer across four upstream pathway domains. A: Lipid pathway; B: Blood pressure pathway; C: Inflammation pathway; D: Nutrition pathway. A-D present system-specific path diagrams illustrating how fructosamine relates to D-dimer through upstream physiological pathway indices and a downstream coagulation substrate node. Each panel displays the full set of modeled pathways within the corresponding domain, including the direct association fructosamine → D-dimer and the indirect pathway fructosamine → upstream pathway index → coagulation (substrate) → D-dimer. Direct paths from each upstream pathway index to D-dimer are also shown, where specified. Panels correspond to four upstream pathway domains: (A) Lipid metabolism; (B) Blood pressure; (C) Inflammation; and (D) Nutrition. Arrow color denotes the direction of the standardized association (red = positive; cyan = negative), arrow thickness reflects the magnitude of the standardized coefficient (|Std.β|; 0-0.1, 0.1-0.3, > 0.3), and line type indicates statistical significance (solid = significant; dashed = not significant). FA: Fructosamine; BP: Blood pressure.

Figure 7 Pathways linking glycated hemoglobin to D-dimer across four upstream pathway domains. A: Lipid pathway; B: Blood pressure pathway; C: Inflammation pathway; D: Nutrition pathway. A-D present system-specific path diagrams illustrating how glycated hemoglobin relates to D-dimer through upstream physiological pathway indices and a downstream coagulation substrate node. Each panel displays the full set of modeled pathways within the corresponding domain, including the direct association glycated hemoglobin → D-dimer and the indirect pathway glycated hemoglobin → upstream pathway index → coagulation (substrate) → D-dimer. Direct paths from each upstream pathway index to D-dimer are also shown, where specified. Panels correspond to four upstream pathway domains: (A) Lipid metabolism; (B) Blood pressure; (C) Inflammation; and (D) Nutrition. Arrow color denotes the direction of the standardized association (red = positive; cyan = negative), arrow thickness reflects the magnitude of the standardized coefficient (|Std.β|; 0-0.1, 0.1-0.3, > 0.3), and line type indicates statistical significance (solid = significant; dashed = not significant). HbA1c: Glycated hemoglobin; BP: Blood pressure.

Table 6 Stability of structural equation modeling structural pathways across diagnostic subgroups.

Path	ALL Std.β (95%CI)	SCZ Std.β (95%CI)	BD Std.β (95%CI)	MDD Std.β (95%CI)
FPG → Lipid index	-0.04 (-0.09 to 0.01)	-0.06 (-0.15 to 0.02)	-0.07 (-0.20 to 0.06)	0.04 (-0.04 to 0.11)
FPG → BP index	0.11 (0.06, 0.16)	0.06 (-0.01 to 0.14)	0.08 (-0.02 to 0.19)	0.10 (0.02, 0.18)
FPG → Inflamm index	0.08 (0.02, 0.15)	0.14 (0.02, 0.26)	0.02 (-0.08 to 0.11)	-0.01 (-0.11 to 0.07)
FPG → Nutrition index	-0.04 (-0.09 to 0.00)	-0.00 (-0.09 to 0.08)	-0.02 (-0.12 to 0.08)	-0.10 (-0.19 to 0.00)
FA → Lipid index	0.23 (0.17, 0.27)	0.30 (0.24, 0.38)	0.17 (0.08, 0.26)	0.18 (0.10, 0.26)
FA → BP index	0.09 (0.05, 0.13)	0.06 (-0.01 to 0.14)	0.11 (0.03, 0.20)	0.06 (-0.00 to 0.13)
FA → Inflamm index	-0.00 (-0.06 to 0.06)	-0.04 (-0.14 to 0.08)	-0.02 (-0.11 to 0.11)	0.03 (-0.07 to 0.15)
FA → Nutrition index	-0.32 (-0.37 to -0.29)	-0.31 (-0.38 to -0.24)	-0.32 (-0.39 to -0.24)	-0.33 (-0.41 to -0.26)
HbA1c → Lipid index	0.06 (0.01, 0.12)	-0.00 (-0.10 to 0.09)	0.16 (0.03, 0.28)	0.06 (-0.03 to 0.16)
HbA1c → BP index	-0.05 (-0.11 to 0.00)	-0.03 (-0.13 to 0.06)	-0.02 (-0.12 to 0.08)	-0.01 (-0.09 to 0.08)
HbA1c → Inflamm index	0.04 (-0.01 to 0.10)	0.00 (-0.10 to 0.13)	0.11 (-0.02 to 0.24)	0.19 (0.10, 0.29)
HbA1c → Nutrition index	0.17 (0.11, 0.23)	0.18 (0.08, 0.28)	0.11 (0.00, 0.20)	0.16 (0.02, 0.26)
Lipid → Coag (substrate)	0.12 (0.08, 0.15)	0.11 (0.05, 0.16)	0.09 (0.03, 0.16)	0.14 (0.08, 0.19)
BP → Coag (substrate)	0.08 (0.05, 0.12)	0.08 (0.03, 0.13)	0.11 (0.03, 0.17)	0.05 (-0.00 to 0.09)
Inflamm → Coag (substrate)	0.42 (0.38, 0.46)	0.40 (0.35, 0.45)	0.41 (0.32, 0.51)	0.45 (0.39, 0.50)
Nutrition → Coag (substrate)	-0.01 (-0.04 to 0.02)	0.03 (-0.03 to 0.09)	-0.02 (-0.08 to 0.05)	-0.05 (-0.10 to 0.00)
Lipid → Coag (time)	0.07 (0.03, 0.11)	0.03 (-0.02 to 0.10)	0.12 (0.03, 0.20)	0.10 (0.05, 0.16)
BP → Coag (time)	0.07 (0.04, 0.11)	0.07 (0.03, 0.14)	0.04 (-0.03 to 0.11)	0.07 (0.02, 0.13)
Inflamm → Coag (time)	-0.02 (-0.06 to 0.01)	0.01 (-0.03 to 0.04)	-0.07 (-0.15 to 0.00)	-0.06 (-0.14 to 0.03)
Nutrition → Coag (time)	-0.01 (-0.07 to 0.04)	0.04 (-0.07 to 0.10)	-0.06 (-0.14 to 0.01)	-0.05 (-0.12 to 0.00)
FPG → D-dimer	0.12 (0.04, 0.18)	0.11 (-0.02 to 0.24)	0.13 (-0.03 to 0.28)	0.09 (0.00, 0.19)
FA → D-dimer	-0.19 (-0.27 to -0.13)	-0.24 (-0.38 to -0.12)	-0.17 (-0.33 to -0.02)	-0.16 (-0.28 to -0.06)
HbA1c → D-dimer	0.03 (-0.04 to 0.11)	0.03 (-0.10 to 0.17)	0.06 (-0.11 to 0.21)	0.06 (-0.05 to 0.16)
Lipid → D-dimer	0.00 (-0.05 to 0.06)	0.01 (-0.08 to 0.11)	-0.11 (-0.23 to 0.01)	0.06 (-0.03 to 0.14)
BP → D-dimer	-0.03 (-0.08 to 0.02)	-0.07 (-0.14 to 0.01)	-0.04 (-0.18 to 0.10)	-0.02 (-0.11 to 0.05)
Inflamm → D-dimer	0.16 (0.09, 0.26)	0.16 (0.06, 0.29)	0.21 (-0.02 to 0.38)	0.11 (-0.03 to 0.35)
Nutrition → D-dimer	0.08 (0.03, 0.13)	0.06 (-0.03 to 0.15)	0.16 (0.03, 0.27)	0.07 (-0.01 to 0.16)
Coag (substrate) → D-dimer	-0.09 (-0.15 to -0.03)	-0.06 (-0.15 to 0.02)	-0.12 (-0.25 to 0.02)	-0.12 (-0.26 to -0.03)
Coag (time) → D-dimer	-0.02 (-0.07 to 0.03)	-0.05 (-0.15 to 0.05)	-0.05 (-0.15 to 0.06)	0.05 (-0.02 to 0.14)

BP: Blood pressure; FPG: Fasting plasma glucose; FA: Fructosamine; HbA1c: Glycated hemoglobin.

Sensitivity analysis

The full mediation model exhibited a slight decrease in fit (CFI = 0.958, RMSEA = 0.057), yet it remained within a reasonable range; the partial mediation model, being a just-identified model with zero degrees of freedom, lacked comparable fit indices. In sensitivity analyses substituting outcomes, pathway definitions, and expanded covariates, the majority of models maintained good fit (most CFIs > 0.98, with RMSEAs of approximately 0.04), indicating that the study conclusions were robust to various operationalizations of inflammation, nutrition, coagulation, and metabolic pathway indicators. Among single-glycemic-indicator models, the FPG model demonstrated the best fit (CFI = 0.997, RMSEA = 0.038), while the FA and albumin-corrected FA single-indicator models showed relatively weaker fit (RMSEA approximately 0.09-0.11), supporting the superiority of joint glycemic indicator modeling over single-indicator models. Overall, the sensitivity analyses did not alter the primary structural relationships, suggesting that the model results possessed robust stability and consistency (Figure 8 and Table 7).

Figure 8 Heatmap of structural pathway stability across the main model and multiple sensitivity analyses. This heatmap summarizes the stability of each retained structural path (rows) across the main model and all subgroup/sensitivity models (columns). For each model-path cell, the standardized path coefficient (β) and its bootstrap 95% confidence interval (CI) were compared with the corresponding coefficient in the main model (baseline). Statistical significance was defined as a CI that does not cross zero: A path was considered significant if the CI lower > 0 or the CI upper < 0; otherwise, it was treated as non-significant or uncertain. Direction was determined from the sign of β (with a near-zero tolerance ε = 1 × 10⁻¹²): Β > ε (positive), β < -ε (negative), and |β| ≤ ε (zero/neutral). Direction consistency was defined as “same direction” when the current sign matched the baseline sign, or when either sign was zero (i.e., baseline = 0 or current = 0 was treated as direction-compatible). A direction flip was flagged when both baseline and current signs were non-zero and opposite (sign_baseline ≠ sign_current). Cells were assigned to four stability classes (tile color): (1) Consistent and significant (stable): Direction compatible and both baseline and current paths significant; (2) Consistent and marginal (partially stable): Direction compatible but statistical significance differed between baseline and current (exactly one significant); (3) Not significant/uncertain: All remaining non-flipped cases, including when both paths were non-significant; and (4) Direction flipped (unstable): Opposite non-zero signs between baseline and current. To quantify magnitude stability, the change in standardized coefficient relative to baseline was computed as Δβ = β_current - β_baseline, and its absolute value |Δβ| was used for visualization. |Δβ| was clipped at 0.20 (DELTA MAX = 0.20). Tile size encodes magnitude stability in a direction-dependent manner: For direction-consistent paths, tile size decreases with increasing |Δβ|, whereas for direction-flipped paths, tile size increases with increasing |Δβ|. Cells with missing coefficients or unavailable paths were not drawn to avoid visual misinterpretation.

Table 7 Model fit indices for the main structural equation modeling and sensitivity analyses¹.

Model	NPAR	df	CFI	TLI	RMSEA	SRMR	RMSEA (CI)	NOBS
Main model	119	6	0.990	0.966	0.041	0.001	0.041 (0.030-0.053)	3783
SCZ subgroup model	119	6	0.988	0.960	0.037	0.001	0.037 (0.019-0.057)	1511
BD subgroup model	119	6	0.983	0.941	0.049	0.004	0.049 (0.025-0.075)	876
MDD subgroup model	119	6	0.989	0.962	0.053	0.003	0.053 (0.035-0.073)	1396
Sensitivity model
Sensitivity partial mediation (direct-coag) model	125	0	1	1	0	0	0.000 (0.000-0.000)	3783
Sensitivity full mediation model	112	13	0.958	0.933	0.057	0.026	0.057 (0.050-0.065)	3783
Sensitivity log(D-dimer) outcome model	120	6	0.949	0.821	0.044	0.141	0.044 (0.035-0.054)	3783
Sensitivity lipid (AIP) model	119	6	0.987	0.955	0.045	0.001	0.045 (0.034-0.057)	3783
Sensitivity inflammation (Winsor) model	119	6	0.987	0.956	0.040	0.001	0.040 (0.030-0.052)	3783
Sensitivity inflammation (SIRI) model	119	6	0.973	0.907	0.049	0.002	0.049 (0.039-0.060)	4284
Sensitivity inflammation (extended) model	119	6	0.986	0.952	0.042	0.001	0.042 (0.031-0.053)	3783
Sensitivity nutrition (CONUT) model	119	6	0.991	0.969	0.038	0.001	0.038 (0.027-0.050)	3783
Sensitivity nutrition (PNI) model	119	6	0.990	0.965	0.04	0.003	0.040 (0.030-0.052)	3783
Sensitivity coag (substrate) model	101	3	0.993	0.966	0.042	0.001	0.042 (0.027-0.059)	3783
Sensitivity coag (time) model	101	3	0.973	0.865	0.041	0.000	0.041 (0.026-0.058)	3783
Sensitivity coag (composite) model	101	3	0.994	0.968	0.051	0.001	0.051 (0.037-0.068)	3783
Sensitivity covars (extended) model	133	6	0.990	0.964	0.042	0.001	0.042 (0.031-0.054)	3783
Sensitivity single-sugar (FPG) model	109	2	0.997	0.973	0.038	0.001	0.038 (0.020-0.059)	3783
Sensitivity single-sugar (FA) model	109	2	0.985	0.839	0.089	0.002	0.089 (0.071-0.108)	3783
Sensitivity single-sugar (FA-ALBcorr) model	109	2	0.981	0.801	0.105	0.003	0.105 (0.086-0.124)	3783
Sensitivity single-sugar (HbA1c) model	109	2	0.995	0.946	0.076	0.002	0.076 (0.058-0.096)	3783
Sensitivity tri-sugar (FA-ALBcorr) model	119	6	0.992	0.972	0.039	0.002	0.039 (0.028-0.051)	3783
Sensitivity gly-composite (FA) model	109	2	0.985	0.842	0.089	0.002	0.089 (0.070-0.108)	3783
Sensitivity gly-composite (FA-ALBcorr) model	109	2	0.984	0.832	0.094	0.002	0.094 (0.075-0.113)	3783

¹This table reports global model fit indices for the main serial SEM, diagnosis-specific subgroup models, and a set of prespecified sensitivity analyses. Sensitivity analyses were designed to assess the robustness of the main mechanistic conclusions under alternative assumptions, including structural constraints (partial vs full mediation), outcome specification (binary vs continuous D-dimer), exposure definitions (single-glycemic indicators, alternative tri-sugar configurations, and glycemic composites), mediator operationalization (lipid, inflammatory, nutritional, and coagulation domains), and covariate extension. Models with zero degrees of freedom are just-identified; therefore, their fit indices (e.g., comparative fit index = 1.00, root mean square error of approximation = 0) are mathematically determined and should not be interpreted as indicators of superior model fit. The number of observations (number of observations) varies across models due to variable-specific missingness, with each model estimated using all available data under its corresponding specification. Sensitivity analyses were not intended to identify an optimal-fitting model, but rather to evaluate the stability and consistency of model fit and structural conclusions across alternative modeling scenarios.

NPAR: Number of free parameters; CFI: Comparative fit index; TLI: Tucker-Lewis index; RMSEA: Root mean square error of approximation; SRMR: Standardized root mean square residual; CI: Confidence interval; NOBS: Number of observations; SCZ: Schizophrenia; BD: Bipolar disorder; MDD: Major depressive disorder; PNI: Prognostic nutritional index; CONUT: Controlling nutritional status; FPG: Fasting plasma glucose; FA: Fructosamine; HbA1c: Glycated hemoglobin; ALB: Albumin; AIP: Atherogenic index of plasma.

DISCUSSION

We presented a highly biologically significant structural finding regarding the direct relationships between glycemic indicators across different time scales and hypercoagulability: FPG acted as a stable procoagulant factor, whereas FA exhibited a significant and consistent protective effect; in contrast, the direct influence of HbA1c was markedly weaker than the former two.

Glycemic indicators across three time scales

In this study, FPG emerged as a stable procoagulant factor. Reflecting short-term, instantaneous glucose levels, elevated FPG in psychiatric inpatients often signifies not only insulin resistance but also acute physiological and psychological stress. Primarily, acute psychiatric stress releases high concentrations of epinephrine and cortisol via the neuroendocrine axis (activating the hypothalamic-pituitary-adrenal axis and sympathetic nervous system[29]). This not only directly enhances platelet sensitivity to thrombin[30] but also inhibits the fibrinolytic system by upregulating plasminogen activator inhibitor-1[31], forming a “hyperglycemia-hypercortisolemia-hypercoagulability” triad. Consequently, FPG is strongly associated with D-dimer elevation within hours to days. Furthermore, transient hyperglycemia exhibits direct vascular toxicity. A hyperglycemic environment can directly induce shedding and damage to the glycocalyx on the vascular endothelial surface, increasing endothelial permeability[32] and exposing subendothelial collagen and tissue factor, thereby initiating the extrinsic coagulation pathway. Additionally, acute hyperglycemia induces mitochondrial superoxide overproduction, and the resulting oxidative stress further activates the nuclear factor-κB pathway, promoting the release of inflammatory cytokines [e.g., interleukin (IL)-6, tumor necrosis factor-α][16]. This “stress-hyperglycemia-inflammation” cascade renders FPG a sensitive indicator for acute-phase hypercoagulability risk.

FA reflects the non-enzymatic glycation of serum proteins (primarily albumin) and its concentration is governed dually by blood glucose levels and the albumin turnover rate[17]. In psychiatric inpatients, malnutrition[33], irregular eating, and chronic inflammatory wasting are prevalent, leading to decreased albumin synthesis[34,35]. Therefore, a low FA level likely indicates not “good glycemic control”, but rather a comprehensive state of “malnutrition-high inflammatory burden-poor protein turnover”. Under such conditions, anticoagulant protein production capacity declines, the endothelium becomes vulnerable[36], and elevated inflammation promotes tissue factor expression and fibrinolysis inhibition[37]. Protein deficiency drops plasma colloid osmotic pressure, triggering increased endothelial permeability, which further facilitates coagulation factor leakage and activation[38]. Hence, in the psychiatric cohort, a “low FA” signifies a depletion of anticoagulant reserves, explaining its inverse relationship with D-dimer. In the present model, HbA1c did not demonstrate a significant direct effect on hypercoagulability during early admission. This suggests that VTE risk during the early phase of psychiatric hospitalization is predominantly driven by acute events (e.g., stress, dehydration, medication adjustments, immobilization)[20], rather than the cumulative effects of chronic hyperglycemia[21].

Four upstream multi-system pathways

Through SEM, we systematically uncovered an integrative framework wherein three glycemic indicators influenced hypercoagulability via four physiological pathways, each with specific emphases and intertwining mechanisms.

Inflammatory pathway: The inflammatory pathway served as the most potent hub linking glycemic exposure to hypercoagulability. The “FPG-inflammation-D-dimer” and “FPG-inflammation-coagulation substrate” pathways constituted the core associative links with the hypercoagulable state. A hyperglycemic environment induces oxidative stress, activating nuclear factor-κB signaling in vascular endothelial cells and monocytes, which prompts aberrant tissue factor expression and directly initiates the extrinsic coagulation cascade[39]. Clinical research confirms that under systemic inflammatory conditions, hyperglycemia significantly enhances thrombin generation (elevating thrombin-antithrombin complexes) and soluble tissue factor levels[39]. Moreover, hyperglycemia can stimulate hepatic IL-6 production via the S100A8/A9-RAGE axis, subsequently promoting thrombopoiesis and leading to inflammatory thrombocytosis[40]. In psychiatric inpatients, SCZ itself is accompanied by chronic low-grade inflammation and microglial activation; metabolic syndrome associated with antipsychotics further amplifies the inflammatory cascade, forging a vicious “hyperglycemia-inflammation-hypercoagulability” positive feedback loop[41].

Nutritional pathway: The nutritional pathway formed a physiological defense against hypercoagulability. In this study, the nutrition index comprised key indicators reflecting protein synthesis capacity and nutritional reserves, such as total protein, prealbumin, and hemoglobin. The decline of these indicators is ubiquitous among psychiatric inpatients[42], especially in those with long disease durations, chronic wasting conditions, and poor dietary intake during hospitalization[33]. FPG, FA, and HbA1c can all influence hypercoagulability via the nutritional pathway, but their directional effects and mechanisms vary. FPG correlated negatively with the nutrition index, reflecting a hypercatabolic state under acute stress that is linked to a reduction in protein reserves. FA exhibited a strong negative correlation with the nutrition index, making it the most sensitive indicator of nutritional status. FA levels are dictated by both glucose and albumin metabolism; in psychiatric inpatients, low FA frequently reflects hypoalbuminemia rather than optimal glycemic control. Research indicates that albumin, the most abundant plasma protein, possesses multiple anticoagulant properties: It maintains vascular endothelial integrity and participates in the synthesis and functional preservation of natural anticoagulants like protein C and protein S[43]. When albumin drops severely, these similarly sized anticoagulant proteins are also lost through impaired barriers (renal, gastrointestinal, etc.), directly compromising endogenous anticoagulant capacity[44]. Furthermore, clinical studies verify that low albumin levels significantly shorten coagulation time, enhance platelet aggregation, and increase maximum clot firmness, directly promoting venous thrombosis[28]. Therefore, a “low FA” in psychiatric patients essentially acts as a comprehensive marker for “malnutrition-hypoalbuminemia-depleted anticoagulant reserves”.

In stark contrast to FA, HbA1c correlated significantly and positively with the nutrition index, where poorer subsequent nutritional status was similarly associated with D-dimer elevation. Within the psychiatric population, poor long-term glycemic control, represented by HbA1c, directly impacts the coagulation system. Hyperglycemia modifies platelet membrane proteins via non-enzymatic glycation, leading to enhanced expression of platelet surface activation markers (e.g., CD62P) and glycoprotein receptors (GPIIb/IIIa), which maintains platelets in a basal hyper-reactive state[45]. Concurrently, HbA1c levels correlate significantly and positively with coagulation factors VIII, IX, and XI[46], accompanied by elevated plasminogen activator inhibitor-1 levels, which suppress the fibrinolytic system[47]. Furthermore, oxidative stress and microangiopathy induced by persistent hyperglycemia insidiously deplete the body’s protein reserves, plunging the patient into “hidden malnutrition”. Hypoalbuminemia directly undermines the synthesis and maintenance of natural anticoagulants (e.g., protein C, protein S)[48]. The cumulative metabolic burden reflected by elevated HbA1c contributes to the alteration of the body’s coagulation-fibrinolysis balance through this very cumulative effect of “chronic hyperglycemia-multisystem damage-hypercoagulable susceptibility”.

Blood pressure pathway: The blood pressure pathway exhibited a complex bidirectional regulation in psychiatric patients. The “FPG-blood pressure-coagulation substrate” pathway demonstrated a relatively weak effect in this study. This is closely tied to the unique pathophysiological states of psychiatric inpatients. Driven by acute psychiatric stress, the overactivation of the hypothalamic-pituitary-adrenal axis and sympathetic nervous system leads to stress-induced hypertension, which inherently induces endothelial dysfunction and diminishes nitric oxide bioavailability, leading to increased von Willebrand factor release; as a key mediator of platelet adhesion, von Willebrand factor directly promotes the activation and aggregation of coagulation substrates (fibrinogen, platelets)[49]. Simultaneously, the “FA-blood pressure-coagulation substrate” pathway suggested that the short-to-medium-term sympathetic stress captured by FA causes abnormal mechanical shear stress, leading to vascular endothelial injury and tissue factor exposure, which initiates the extrinsic coagulation cascade and recruits platelets[50]. Notably, HbA1c showed a weak negative correlation with blood pressure, likely reflecting that in patients with severe psychiatric disorders over long disease courses, antipsychotics like clozapine can cause orthostatic hypotension and increased blood pressure variability via α1-adrenergic receptor blockade[41]. Thus, diverse time-scale glucose metabolism indicators exhibit specific heterogeneities in their hemodynamic effects. The driving effect of blood pressure changes on coagulation substrates (fibrinogen and platelets) is masked by medication effects and autonomic dysfunction in psychiatric populations, ultimately presenting as a weak predictive role for D-dimer.

Lipid pathway: The “FA/HbA1c - lipid - coagulation substrate - D-dimer” pathway demonstrated a moderate effect in this study. Studies show that variations in FA levels correlate significantly with lipid profile alterations, impacting lipoprotein structure and function via non-enzymatic glycation to promote atherosclerotic plaque formation[51]. Glycated LDL cholesterol is more easily internalized by macrophages to form foam cells, while glycated HDL loses its anti-atherogenic properties[52]. In psychiatric patients, the short-to-medium-term nutritional and metabolic fluctuations reflected by FA capture the hemodynamic and lipid deterioration brought by acute stress more acutely than HbA1c. While hyperlipidemia is typically a cardiovascular risk factor in the general population, psychiatric patients, particularly those on long-term antipsychotics, often present with “metabolic syndrome-like” changes, including high triglycerides, low HDL, fatty liver, and insulin resistance[8,53]. However, a “lipid paradox” may exist among acutely ill, severe psychiatric patients[54]. In severe mental illness, low cholesterol levels are often markers of malnutrition, cell membrane breakdown, or chronic infectious wasting, rather than indicators of cardiovascular health[55]. Therefore, in our model, the inverse trend presented by the lipid pathway may not denote classical low-lipid protection, but rather a “pseudo-low-lipid” state driven by long-term malnutrition and protein depletion. In other words, absolute lipid levels cannot serve as stable predictors of hypercoagulable risk in psychiatric inpatients, as they reflect a complex background shaped by countermeasures against nutritional wasting[55], inflammatory activation[56], drug metabolic effects[57], and overall health status, rather than straightforward atherogenesis or thrombogenesis risk.

Downstream coagulation layer

Conventionally, elevated fibrinogen and platelets indicate thrombotic risk[58]. However, our data unveiled a contradictory dynamic: Coagulation substrates-whether amplified via lipid or blood pressure pathways-ultimately exhibited a highly consistent, significant negative correlation with the outcome indicator, D-dimer. This phenomenon profoundly elucidates the “high-turnover” consumptive nature of hypercoagulability in psychiatric contexts. Driven persistently by lipotoxicity and endothelial injury, substantial amounts of fibrinogen are rapidly converted into cross-linked fibrin networks, sparking consumptive microthrombosis of platelets; the ensuing secondary hyperfibrinolysis drastically spikes D-dimer levels. Consequently, the relative decline in substrate levels in our cross-sectional data essentially represented a snapshot of compensatory substrate overdraft following systemic coagulation overactivation. This mechanistic chain suggests that VTE risk in psychiatric patients is not merely a metabolic complication, but a pathophysiological process-triggered synergistically by glucolipotoxicity and vascular stress-that ultimately progresses toward a non-overt disseminated intravascular coagulation-like state[59].

Subgroup analysis

Subgroup analysis revealed significant pathophysiological heterogeneity in hypercoagulability mechanisms across different psychiatric disorders, providing a rationale for personalized interventions. Patients with SCZ exhibited a classical “inflammation-metabolism-coagulation” direct-drive profile. In this group, the transduction effect of FPG via the inflammatory pathway was the strongest. SCZ is now recognized to possess a significant immune-inflammatory dimension, with patients often displaying cytokine network dysregulation (e.g., IL-6, TGF-β) and microglial activation[60]. This endogenous immune activation forms the basis of “immunothrombosis,” where immune cells (monocytes, neutrophils) directly initiate coagulation by releasing tissue factor and neutrophil extracellular traps[61]. Moreover, antipsychotic (e.g., clozapine)-induced metabolic syndrome exacerbates this inflammatory cascade, forging a self-amplifying procoagulant loop[62].

Conversely, patients with BD showcased a “nutrition-fluctuation dominant” mechanistic pattern. Here, the effects of FA and the nutritional pathway were paramount. BD’s disease course centers on polar shifts between mania and depression; this emotional instability is accompanied by marked eating behavior disruptions (e.g., reduced intake during mania and disordered eating during depression)[63] and dramatic energy metabolism oscillations. Particularly during manic episodes, patients experience extreme psychomotor agitation and sleep deprivation[64], plunging the body into severe hypercatabolism and oxidative stress[65]. This rapidly consumes or accelerates the turnover of albumin and endogenous anticoagulant proteins, inducing negative nitrogen balance. As a short-to-medium-term indicator of protein synthesis and glucose metabolism over 2-3 weeks[66], FA accurately quantifies this “nutritional shock” and the concomitant depletion of physiological reserves. This nutritional reserve deficit (low albumin/Low FA) directly undermines the body’s physiological anticoagulant defenses[67], resulting in a “low FA-high D-dimer” hypercoagulable phenotype. Thus, clinical management for BD must simultaneously fortify nutritional support alongside mood stabilization to correct negative nitrogen balance, which is crucial for restoring endogenous anticoagulant capacity and preventing immunothrombosis.

Patients with MDD exhibited an “aging-comorbidity dominant” hypercoagulability mechanism. This group had the oldest average age and the highest model explanatory power. Their hypercoagulable risk primarily stems from pathologies linked to the “vascular depression” hypothesis: Long-term vascular aging, arteriosclerosis, and microvascular rarefaction inherently prime the vascular wall into a pro-inflammatory and procoagulant state[68]. Hypercoagulability here is not merely the outcome of acute stress but the cumulative manifestation of chronic vascular pathology. Furthermore, depression shares unique biological ties with platelet activation; research indicates increased platelet activation and enhanced reactivity in MDD patients[69]. The specific behavior of the coagulation substrate node (including platelets) within the MDD model may precisely mirror this unique platelet-driven thrombotic risk.

Sensitivity analysis

To ensure the reliability of these mechanistic models and eliminate confounding from statistical artifacts or model specifications, we conducted systematic sensitivity analyses[70]. Initially, structural robustness tests demonstrated a significant decline in goodness-of-fit when we attempted to simplify the glycemic exposure module (e.g., retaining only a single glycemic indicator), confirming the necessity and independence of utilizing FPG, FA, and HbA1c across different time scales to explain hypercoagulable variance. Second, in the outcome specification sensitivity analysis, replacing the binary outcome (D-dimer > 0.5) with a continuous log-transformed variable (log D-dimer) using maximum likelihood estimation showed that the direction and significance of core path coefficients (particularly the negative path of coagulation substrates and the positive path of inflammation) remained highly consistent. Furthermore, after we processed extreme outliers via Winsorization, key conclusions held firm, indicating that the findings were not driven by a minority of extreme cases but rather reflect universal pathophysiological traits of the psychiatric population. Lastly, the new introduction of the albumin-corrected FA (FA-ALBcorr) maintained the significant protective role of FA against hypercoagulability, further discarding hypoalbuminemia-related confounding and confirming the true value of FA as a metabolic-nutritional homeostatic marker. These multidimensional validations validate the proposed “glycemia-inflammation/nutrition-hypercoagulability” multisystem coupling mechanism and lend high statistical robustness and biological plausibility to it.

To sum up, the hypercoagulable state in psychiatric inpatients is not a single “diabetic complication”, but a complex network in which acute stress (elevated FPG) and nutritional depletion (decreased FA), as well as immune activation (inflammatory pathway). FPG represents the “fight-or-flight” stress state, FA is the “barometer” of nutritional and anticoagulant reserves, and inflammation is the key player of this cascade. In psychiatric patients, the use of glycemic indices alone to determine the risk of VTE should be abandoned, and a system of risk assessment should be developed that includes nutrition status, inflammatory status, and disease mechanisms, with an emphasis on early prevention measures for the “low FA and high inflammation” patient group who are high-risk but not known to doctors.

Limitations

There are a number of limitations in this study. First, only psychiatric inpatients who underwent all tests for FPG, FA, and HbA1c were included, which could potentially lead to selection bias, as the group studied may not be representative of the general population of psychiatric inpatients, and may represent those with a higher metabolic risk or may be under more intensive medical monitoring, limiting the external validity of the results. Second, the cross-sectional design does not provide time-series analysis, and therefore, the pattern of the relationships among paths as shown in the SEM can only be a potential association pattern and cannot strictly be used to infer the causal relationships. Due to the retrospective, real-world nature of the study, systematic collection of the detailed, often incomplete, pre-admission medication histories was not possible, including detailed cumulative dosing, frequent changes of agents, and polypharmacy. Thus, the observed associations may therefore be a residual result of the unmeasured effects of the medications.

CONCLUSION

We developed and tested in a large real-world cohort of psychiatric inpatients a structural equation model that outlines the “multi-time-scale glycemic exposure-multisystem pathways-coagulation nodes-hypercoagulable state” axis. Our findings revealed that, after the adjustments for the various covariates, the association between glycemia and hypercoagulability risk was not a simple linear model: FPG had a steady positive association with hypercoagulability risk, FA had a strong inverse association, and the direct association of HbA1c remained modest. The inflammatory pathway was also identified as the core pathway linking glycemic exposure to hypercoagulable outcomes that was directly correlated with the hypercoagulable state and was strongly associated with downstream coagulation-endothelial nodes. In contrast, the coagulation substrate and coagulation time nodes were not positive mediators that would represent a driving force to increase the level of D-dimer. Overall, the multipath approach proved to be very stable when applied to different psychiatric diagnostic groups and other model configurations. It provides new, comprehensive evidence for understanding the metabolic dysfunction and hypercoagulability risk in psychiatric inpatients and suggests that glycemic metabolic markers have significant reference value in the clinical evaluation of hypercoagulability risk.

ACKNOWLEDGEMENTS

The authors would like to thank the clinical and nursing staff at Hangzhou Seventh People’s Hospital for their meticulous documentation of the clinical and laboratory data used in this study. We also extend our profound appreciation to all the patients included in this retrospective cohort, as their clinical information is the foundation of this real-world research.