Gradient boosting machine model predicts psychiatric complications after deep brain stimulation in Parkinson’s disease

Abstract

BACKGROUND

Deep brain stimulation (DBS) is an effective method for treating the motor symptoms of advanced Parkinson’s disease (PD). However, complications such as depression, anxiety, cognitive impairment, and delirium that occur after DBS surgery can severely affect a patient’s quality of life and the therapeutic effect.

AIM

To construct a gradient boosting machine (GBM) risk model to predict the risk of mental complications such as depression, anxiety, and cognitive impairment in patients with PD after DBS surgery.

METHODS

We retrospectively collected data on patients with PD treated at a top-tier hospital in China between June 2023 and December 2024. During this period, 234 cases were screened and analyzed, of which 70% were included in the modeling set and the remaining 30% in the test set. The modeling set was used to construct the risk prediction model, whereas the test set was used to validate the predictive performance of the model. Additionally, we used the GBM model to predict outcomes for 65 patients with PD who visited the hospital between January 2025 and April 2025, and analyzed the application effect of the model.

RESULTS

In a cohort of 234 patients undergoing DBS, the incidence of psychiatric complications such as depression, anxiety, and cognitive impairment was 37.61%. Age, surgery duration, fasting time, family relationship health assessment scale scores, and unified PD rating scale III scores were identified as independent influencing factors. Based on these variables, the constructed GBM model demonstrated excellent predictive performance, with an accuracy of 80.0%, sensitivity of 95.7%, and specificity of 78.6%. Decision curve analysis revealed that the model’s clinical benefit and applicability are optimal when the threshold is between 0.09 and 0.70.

CONCLUSION

The prediction model constructed based on the GBM algorithm has good predictive performance and can provide a reference for clinical medical staff to identify groups at high risk for mental complications such as depression, anxiety, cognitive impairment, and delirium after DBS.

Key Words: Parkinson’s disease; Deep brain stimulation; Postoperative complications; Gradient boosting machine; Prediction model

Core Tip: This study developed a gradient boosting machine (GBM) model to predict psychiatric complications (depression, anxiety, cognitive impairment, and delirium) in patients with Parkinson’s disease after deep brain stimulation surgery. By analyzing data from 234 patients, the model identified five critical risk factors: Age, surgery duration, fasting time, Family Relationship Health Assessment Scale score, and motor symptom severity (Unified Parkinson’s Disease Rating Scale Part III score). These factors collectively explained 37.6% complication incidence. The GBM model achieved high predictive accuracy (80.0%), sensitivity (95.7%), and area under the curve (0.896) in external validation (65 patients). Decision curve analysis confirmed the optimal clinical utility for risk thresholds between 0.09-0.70, enabling preoperative risk stratification and personalized interventions to mitigate postoperative neuropsychiatric risks.

INTRODUCTION

Parkinson’s disease (PD) is a neurodegenerative disorder. Epidemiological data show that the prevalence of the disease is approximately 0.3% in the general population[1]; however, it increases significantly to 1%-2% among the older population aged 65 years and older[2]. In recent years, with accelerated aging in the population, the PD disease burden has shown a rapid upward trend. Statistics indicate that since the 1990s, the number of patients with PD has surged from 2.5 million to 6.1 million worldwide[3]. It is predicted that by 2030, China will have the highest burden of PD, with the number of patients possibly accounting for half of the global total[4]. This trend is closely related to population aging and extended life expectancy.

Currently, the treatment of PD mainly relies on medication and surgical interventions, aiming to delay disease progression rather than curing it. Deep brain stimulation (DBS) has become a major surgical approach owing to its advantages such as minimal invasiveness, reversibility, and adjustable characteristics[5,6]. This technique precisely locates specific brain areas using stereotactic guidance and implant-stimulating electrodes to regulate abnormal neural activity using electrical pulses, thereby significantly improving motor symptoms. Long-term follow-up studies, both domestically and internationally, have confirmed that DBS can effectively improve motor function and quality of life in patients with mid-to-late-stage PD, particularly in controlling symptoms such as tremors and rigidity[7,8].

However, the impact of DBS on psychiatric complications such as anxiety, depression, delirium, and cognitive dysfunction remains inconclusive. Existing evidence indicates that 10%-36% of patients with PD experience cognitive dysfunction in the early postoperative period[9]. Meta-analyses suggest significant variability in the effects of DBS on anxiety symptoms across different study cohorts, with this heterogeneity potentially attributable to baseline demographic characteristics and preoperative clinical parameters[10,11]. The gradient boosting machine (GBM) algorithm, which iteratively optimizes the integration of multimodal clinical data, has demonstrated advantages in high-dimensional feature modeling and has been validated in other medical fields[12,13]. We proposed the construction of a GBM-based predictive model for postoperative psychiatric complications following DBS to address the current limitations of subjective clinical assessments. This model could provide a reference for preoperative risk stratification and personalized postoperative management, thereby advancing precision medicine in DBS treatment for patients with PD.

MATERIALS AND METHODS

General information

A retrospective observational study involving patients with PD who underwent DBS between June 2023 and December 2024 was conducted at a tertiary hospital in China. The patients were screened according to the following inclusion and exclusion criteria.

Inclusion criteria: (1) Age > 18 years, diagnosis of primary PD based on diagnostic criteria[11], Hoehn-Yahr stages 2-5; (2) Indication for DBS treatment, with bilateral electrode implantation and completed therapy; (3) Normal coagulation function; (4) Normal language, hearing, and communication abilities, capable of completing questionnaires; and (5) Complete clinical and follow-up data.

Exclusion criteria: (1) Patients with secondary Parkinsonism caused by primary brain diseases (e.g., cerebral hemorrhage, encephalitis); (2) Intracranial tumors; (3) Those undergoing only battery replacement or electrode adjustment procedures; (4) Comorbid essential tremor; (5) Immunodeficiency disorders; (6) Cerebral atrophy or Parkinson-plus syndromes; and (7) Severe psychiatric disorders (e.g., schizophrenia, bipolar disorder).

Information collection

Two independent researchers retrieved and analyzed relevant literature materials through the system to determine the materials that needed to be collected, including: (1) General clinical data: Age, body mass index, hypertension, diabetes, education level, swallowing ability, and Hoehn-Yahr stage; (2) Perioperative indicators: Neutrophils, lymphocytes, hemoglobin, white blood cells, serum albumin, C-reactive protein, operation time, anesthesia time, intraoperative blood loss, fasting time, and postoperative intracranial gas volume; and (3) Relevance scale: The Pittsburgh Sleep Quality Index (PSQI) was used to assess sleep quality, with scores ranging from 0 to 21. A total score of seven or higher indicated the presence of sleep problems, with higher scores corresponding to poorer sleep quality. Family relationship health was evaluated using the Family Relationship Health Assessment Scale (H-FRAT), which has a score range of 25 to 50, with higher scores reflecting healthier family dynamics. The quality of life in patients is measured using the 39-item Parkinson’s Disease Questionnaire (PDQ-39), which has a maximum score of 156, with higher scores indicating worse quality of life. Motor symptoms were assessed using the Unified Parkinson’s Disease Rating Scale Part III (UPDRS-III), with scores ranging from 0 to 56 and higher scores indicating more severe motor impairments. Cognitive function was evaluated using the Mini-Mental State Examination (MMSE), with scores ranging from 0 to 30 and higher scores indicating better cognitive performance. Psychological status was assessed using the Hamilton Anxiety Scale (HAMA) and Depression Scale (HAMD). HAMA scores range from 0 to 56, with higher scores reflecting greater anxiety severity, whereas HAMD scores range from 0 to 68, with higher scores indicating more severe depressive symptoms.

Observation indicators and group

The outcome indicator of this study was the occurrence of mental complications such as anxiety, depression, cognitive dysfunction, and delirium after DBS surgery. Patients with any of the above conditions were considered to have postoperative complications and were included in the complication group, whereas the remaining patients were included in the non-complication group.

The study defined postoperative neuropsychiatric complications using the following criteria: (1) Within 3 days after surgery, patients showing new-onset or worsened depression/anxiety states, as indicated by increased HAMD or HAMA severity ratings [e.g., progression from “possible anxiety” (HAMA score 7-13) to “definite anxiety” (HAMA score ≥ 14), or from “normal” (HAMD score < 8) to “mild depression” (HAMD score 9-20)]; (2) Within 3 postoperative days, patients demonstrating cognitive decline based on MMSE rating reductions [e.g., transitioning from “normal/mild cognitive impairment” (MMSE > 18) to “moderate-severe cognitive impairment” (MMSE ≤ 18)]; and (3) Presence of delirium as identified by a 3-minute diagnostic interview for confusion assessment method-defined delirium, an assessment tool adapted from the confusion assessment method, that combines structured patient interviews and behavioral observations to evaluate delirium status. All assessments were conducted by trained clinicians using standardized protocols.

Construction and verification of the GBM model

The GBM model is constructed using the ‘gbm’ package in R. In the ‘gbm’ package, decision trees are used as base learners. The key parameters were as follows: The response variable (distribution) was binary (Bernoulli), the number of boosting trees was set to 2000, the learning rate (shrinkage) was 0.01, and a 10-fold cross-validation (cross-validation folds = 10) was used to select the optimal number of boosting trees.

Statistical analysis

Data storage and management were performed using Excel 2016, while statistical analyses were conducted using SPSS 23.0 and R version 4.0.3. Continuous variables were expressed as mean ± SD and compared using t-tests. Categorical variables were presented as proportions (%) and analyzed using χ² tests or Fisher’s exact test, as appropriate. The least absolute contraction and selection operator (LASSO) method was used to reduce the dimensions of the data, and logistic regression analysis was used to identify risk factors. The predictive performance of the model was validated in the modelling and test sets using receiver operating characteristic (ROC) and calibration curves, respectively. Decision curve analysis (DCA) was used to evaluate the clinical utility of the model. Statistical significance was defined as a two-sided P value < 0.05.

RESULTS

Research process

From January 2023 to December 2024, we screened 234 cases for the construction and training of the gradient lift prediction model. From January 2025 to April 2025, we collected data from additional 65 cases to evaluate the applicability of the model. Figure 1 illustrates the research plan in the form of a flowchart.

Figure 1 Research flowchart.

The postoperative conditions of patients with PD after DBS

In a dataset of 234 internal cases, 88 patients (37.61%) with PD experienced mental complications within 3 days after DBS. Among them, 27 (30.68%) experienced depression or an exacerbation of depression,18 cases (20.45%) experienced anxiety or an exacerbation of anxiety, 15 cases (17.05%) had a decline in cognitive function rating, and 28 (31.82%) showed symptoms of delirium. In the modeling set, 69 patients (42.33%) experienced postoperative mental complications, whereas 94 did not. In the testing set, 19 cases (26.76%) experienced postoperative mental complications, while 52 did not.

Comparison of general clinical data between the complication and non-complication group

In Table 1, compared with those of patients in the non-complication group, patients in the complication group were older and had a higher proportion of hypertension and diabetes history (P < 0.05).

Table 1 Comparison of general clinical data, mean ± SD/n (%).

Items	Complication group (69 cases)	Non-complication group (94 cases)	t/χ2 value	P value
Onset age (year)	56.28 ± 6.99	53.96 ± 7.10	2.073	0.040
Sex			1.184	0.281
Man	36 (52.17)	57 (60.64)
Woman	33 (47.83)	37 (39.36)
BMI (kg/m²)	22.47 ± 1.98	23.02 ± 2.19	1.662	0.099
The course of Parkinson’s disease	8.01 ± 3.56	7.23 ± 2.97	1.521	0.130
Education			0.549	0.459
Junior high school and below	59 (85.51)	84 (89.36)
High school and above	10 (14.49)	10 (10.64)
Hypertension history	35 (50.72)	26 (27.66)	9.039	0.003
Type 2 diabetes mellitus	33 (47.83)	24 (25.53)	8.697	0.003
Swallowing disorder	28 (40.58)	26 (27.66)	2.998	0.083
H-Y stage			1.785	0.074
3	52 (75.36)	81 (86.17)
4	16 (23.19)	13 (13.83)
5	1 (1.45)	0 (0.00)

H-Y: Hoehn-Yahr; BMI: Body mass index.

Comparison of perioperative related indicators between the complication and non-complication group

As shown in Table 2, compared with those of patients in the non-complication group, patients in the complication group had a longer operation time, longer food deprivation time, and higher intracranial gas volume after surgery (P < 0.05).

Table 2 Comparison of perioperative-related indicators, mean ± SD.

Items	Complication group (69 cases)	Non-complication group (94 cases)	t value	P value
NLR	2.41 ± 1.79	2.58 ± 2.06	0.542	0.589
Hemoglobin (g/L)	136.62 ± 15.05	138.85 ± 12.72	1.022	0.308
Albumin (g/L)	42.73 ± 6.75	42.35 ± 5.79	0.389	0.698
CRP (mg/L)	38.71 ± 12.61	39.37 ± 14.07	0.310	0.757
Operation time (minute)	267.36 ± 32.94	245.63 ± 32.57	4.189	< 0.001
Anesthesia time (minute)	288.61 ± 57.79	282.51 ± 51.46	0.709	0.479
Intraoperative blood loss (mL)	54.06 ± 13.43	53.40 ± 18.87	0.246	0.806
Fasting time (hour)	8.87 ± 0.87	7.84 ± 0.88	7.423	< 0.001
postoperative intracranial gas volume (mm3)	9.90 ± 3.46	7.42 ± 2.74	4.912	< 0.001

NLR: Neutrophil-to-lymphocyte ratio; CRP: C-reactive protein.

Comparison of preoperative related scale scores between the complication and non-complication group

As shown in Table 3, the PSQI, PDQ-39, and UPDRS III scores of patients in the complication group were all higher than that of those in the non-complication group, whereas the H-FRAT score was lower than that in the non-complication group (P < 0.05).

Table 3 Comparison of preoperative-related scale scores, mean ± SD.

Items	Complication group (69 cases)	Non-complication group (94 cases)	t value	P value
PSQI score	10.57 ± 1.98	8.62 ± 2.32	5.630	< 0.001
H-FRAT score	28.94 ± 3.15	32.86 ± 4.81	6.274	< 0.001
PDQ-39 score	74.81 ± 8.97	70.78 ± 11.42	2.434	0.016
UPDRS III score	42.65 ± 5.00	38.09 ± 4.65	6.002	< 0.001

PSQI: Pittsburgh Sleep Quality Index; H-FRAT: Family Relationship Health Assessment Scale; PDQ-39: 39-item Parkinson’s Disease Questionnaire; UPDRS III: Unified Parkinson’s Disease Rating Scale Part III.

LASSO regression analysis

LASSO regression analysis was performed on the eight variables with statistically significant differences, as shown in Tables 1, 2, and 3. The final parameter λ was determined through cross-validation, and the λ value corresponding to the maximum value within one standard error of the minimum mean squared error was chosen as the optimal parameter. In this study, the Lambda.lse was 0.062. As shown in Figure 2, seven independent variables (age, history of hypertension, operation time, fasting time, PSQI score, H-FRAT score, and UPDRS III scores) were selected from the 10 variables to construct the GBM model.

Figure 2 Least absolute contraction and selection operator regression screening of independent variables. A: Coefficient path diagram; B: Cross-validation curve.

Multiple logistic regression analysis

Multivariate logistic regression analysis was performed on the aforementioned seven variables, and five variables with statistical significance (P < 0.05) were selected: Age, operation time, fasting time, H-FRAT score, and UPDRS III score (Figure 3).

Figure 3 Forest plot of the results of multivariate analysis. PSQI: Pittsburgh Sleep Quality Index; H-FRAT: Family Relationship Health Assessment Scale; UPDRS III: Unified Parkinson’s Disease Rating Scale Part III.

GBM model

The GBM model ranked the important variables as follows: Fasting time (26.02 points), operation time (25.61 points), H-FRAT score (18.27 points), UPDRS III score (17.37 points) and age (12.73 points), as shown in Figure 4.

Figure 4 The scores of important variables. H-FRAT: Family Relationship Health Assessment Scale; UPDRS III: Unified Parkinson’s Disease Rating Scale Part III.

Verification of the GBM model

The ROC curve showed the area under the curve (AUC) for the GBM model in predicting patient prognosis, with the modeling set at 0.962 [95% confidence interval (CI): 0.935-0.988], the test set at 0.917 (95%CI: 0.828-1.000), see Figure 5. The prediction accuracies of the modeling and test sets were 90.2% and 87.3%, respectively. The calibration curve showed good consistency between the values predicted by the GBM model and the actual observed values, indicating that the model can predict the actual probabilities relatively well, as shown in Figure 6. The Brier score for the modeling set was 0.081 and that for the test set was 0.108.

Figure 5 Receiver operating characteristic curve. A: Modeling set; B: Test set.

Figure 6 Calibration curve. A: Modeling set; B: Test set. ROC: Receiver operating characteristic; Dxy: Somers’ Dxy rank correlation; D: Discrimination index; U: Unreliability index; Q: Quality index; Eavg: Average error; S:z: Z value of the z test; S:p: P value of the z test.

The application effect of the GBM model

A total of 65 patients with PD using the model for diagnosis were predicted. The ROC curve showed that the accuracy of the prediction was 80.0%, the sensitivity was 95.7%, the specificity was 78.6%, and the AUC was 0.896 (95%CI: 0.818-0.978), as shown in Figure 7A. The calibration curve showed good consistency between the predicted values of the GBM and the actual observed values. The model can predict the actual probability well, and the Brier value was 0.081, as shown in Figure 7B. The DCA curve showed that the model had the best clinical benefits and applicability when the threshold was 0.09 and 0.70, as shown in Figure 8.

Figure 7 Application effect of the gradient boosting machine prediction model. A: Receiver operating characteristic; B: Calibration curve. ROC: Receiver operating characteristic; Dxy: Somers’ Dxy rank correlation; D: Discrimination index; U: Unreliability index; Q: Quality index; Eavg: Average error; S:z: Z value of the z test; S:p: P value of the z test.

Figure 8 Decision curve. GBM: Gradient boosting machine.

DISCUSSION

With increasing health awareness and advances in medical technology, the diagnosis and treatment of PD have significantly improved. As a therapy that directly modulates the neural nuclear activity to alleviate PD symptoms, DBS has become a pivotal treatment because of its minimally invasive, reversible, and adjustable nature[14,15]. However, the impact of DBS on psychiatric symptoms remains controversial, and some studies suggest that DBS improves anxiety and depression[16,17], while Birchall et al[18] reported potential worsening of these symptoms, in contrast to the findings of Zhang et al[19]. Currently, research on the risk factors for postoperative psychiatric complications, such as depression, anxiety, cognitive impairment, and delirium, remains limited. Our study revealed a complication rate of 37.61%, highlighting the elevated risk of emotional and cognitive dysfunction post-DBS, with potential key influencers including age at onset, surgical duration, fasting time, H-FRAT score, and UPDRS III score.

Our research shows that neuropsychiatric complications such as depression, anxiety, cognitive impairment, and delirium following DBS surgery are more prevalent in older patients, a finding consistent with those of multiple previous studies[20-22]. This age-dependent phenomenon may stem from multifaceted neurobiological alterations. Aging is associated with neurodegenerative changes, including neuronal loss in key brain regions, such as the frontal, temporal, and parietal lobes[23]. Second, the total number of neurons declines with age, accompanied by reduced neurotransmitter levels (e.g., acetylcholine and dopamine) and increased activity of degradation enzymes[24]. This dual effect leads to a significantly lower neurotransmitter availability in older patients, thereby increasing the risk of cognitive impairment. Furthermore, age-related decline in immune function and metabolic capacity prolong postoperative recovery, further increasing susceptibility to psychiatric complications. Notably, the aging nervous system exhibits a heightened sensitivity to external stimuli (e.g., surgical trauma and anesthesia), and such hyperreactivity predisposes patients to emotional instability and cognitive decline. In particular, subthalamic nucleus-DBS may exacerbate pre-existing cortico-subcortical functional disconnections in older patients by disrupting beta-band oscillations in the basal ganglia-prefrontal circuitry, which likely underlies their more pronounced postoperative executive dysfunction[25].

Two critical perioperative factors, the surgical duration and fasting time, warrant special attention. Prolonged surgical time not only increases the physiological burden but may also exacerbate postoperative cognitive impairment and emotional instability due to cerebral overstimulation. Reportedly, extended anesthesia disrupts neurotransmitter balance, significantly elevating the risk of delirium and anxiety symptoms[26]. Regarding fasting management, although enhanced recovery after surgery (ERAS) guidelines recommend 6-8 hours of preoperative fasting, clinical realities often lead to extended fasting periods. Such prolongation triggers glucose fluctuations and electrolyte imbalances, which directly impair neuronal energy metabolism. Evidence shows that patients who fast for > 6 hours face a 9.6-fold higher delirium risk[27]. Moreover, prolonged fasting-induced hunger and thirst markedly increase anxiety levels[28]. Thus, optimizing perioperative fasting protocols combined with ERAS measures such as carbohydrate loading can effectively reduce postoperative neuropsychiatric complications.

An important finding of this study was the significant negative correlation between the H-FRAT scores (reflecting family relationship quality) and the risk of postoperative psychiatric complications. The data showed that patients with higher H-FRAT scores had significantly lower incidences of postoperative depression and anxiety. This association can be explained through three mechanisms. First, healthy family relationships provide crucial emotional support that helps maintain psychological stability during recovery[29]. Second, family support mitigates stress responses and reduces anxiety/depression symptoms[30]. Most importantly, close family bonds facilitate better adaptation to the postoperative physical and psychological changes[31]. When patients feel understood and cared for by family members, their emotional regulation and psychological resilience are significantly enhanced. This finding has important clinical implications, suggesting that family support system evaluations should be incorporated into preoperative assessments, and targeted family-based interventions should be considered to optimize surgical outcomes.

Furthermore, this study confirmed preoperative UPDRS III scores as an independent risk factor for postoperative neuropsychiatric complications. As the gold standard for assessing the severity of motor symptoms in Parkinson’s, elevated UPDRS III scores indicate significant premotor dysfunction. This impairment affects mental health through two pathways. Directly, severe motor limitations reduce daily activity capacity and quality of life, thereby predisposing patients to depressive and anxious symptoms[32]. Indirectly, patients with poor baseline motor function often develop pessimistic expectations regarding surgical outcomes, which amplify anxiety during postoperative recovery. Notably, these patients undergo surgery with greater psychological vulnerability, which makes them more susceptible to emotional instability and cognitive decline when faced with surgical stress.

Machine learning (ML) technologies have demonstrated significant advantages in disease prediction by uncovering hidden patterns within vast clinical datasets, enabling the precise identification of high-risk patient populations and supporting targeted clinical interventions. ML has achieved groundbreaking application in DBS for PD, advancing early diagnosis[32], treatment response prediction[33], and personalized medicine[34]. Watts et al[34] developed a deep reinforcement learning (DRL)-based model that leveraged real-time motor fluctuation data from wearable sensors to predict the optimal medication timing and dosage. This model generates dynamic treatment plans that maximize symptom relief. The results demonstrated that the DRL-derived regimen significantly outperformed the traditional static treatment plans, highlighting its potential to enhance medical decision-making, particularly in chronic disease management. In other words, real-time sensor data enable optimized clinical assessments and adaptive treatment strategies, allowing for personalized and responsive care that improves therapeutic outcomes. In this study, we constructed a GBM-based predictive model to assess the risk of postoperative complications in 65 patients with PD who underwent DBS. The model demonstrated exceptional performance, achieving an accuracy of 80.0%, a sensitivity of 95.7%, and a specificity of 78.6%. These metrics confirm their robust discriminative ability. The DCA further validated its clinical utility, revealing an optimal net benefit when the risk threshold was set between 0.09 and 0.70. This study underscores the potential of ML-driven predictive tools for refining risk stratification and optimizing the perioperative management of DBS.

In clinical practice, when patients with PD are treated with DBS, the GBM model constructed in this study can be applied in the following steps to predict the risk of postoperative mental complications and to formulate management strategies. First, the preoperative assessment collects key variables such as the patient’s age, operation time, fasting time, H-FRAT score, and UPDRS III score, and inputs them into the GBM model to calculate the risk score. For high-risk patients (such as those with advanced age, long operation time, long fasting time, low H-FRAT score, and high UPDRS III score), it is recommended to provide family relationship counseling before the operation, optimize perioperative management, and shorten the operation and fasting times. Second, postoperative monitoring: Increase the monitoring frequency for high-risk patients after the operation; Assess HAMA, HAMD, and MMSE scores daily; and promptly identify and handle mental complications. Third, the decision support integrates the GBM model into the hospital information system, develops a clinical decision support system, automatically assesses risk, and provides suggestions before the operation. Medical staff can formulate personalized plans based on this information and carefully explain the prediction results to patients and their families. Through these steps, the GBM model can provide personalized preoperative risk assessments and postoperative intervention suggestions for patients with PD, thereby improving treatment outcomes and quality of life.

Research limitations

This study has some methodological limitations: (1) The retrospective study design may lead to information bias during the data collection process, particularly regarding the completeness and accuracy of clinical indicators. This finding needs to be validated through well-designed prospective cohort studies; (2) The model was constructed based solely on data from a single tertiary medical center. Although the internal validation demonstrated good performance, the lack of multi-center external validation data means that the model’s applicability, particularly in primary hospitals with different medical resource configurations, still requires further evaluation; and (3) Although the GBM model could handle multiple predictors, it did not explicitly explore the interactions between them. To enhance the model performance and provide more in-depth clinical insights, we plan to introduce interaction terms, such as the interaction between fasting time and age, in future analyses to assess their impact on model performance. These limitations suggest that the results should be interpreted with caution in clinical practice.

CONCLUSION

Patients with PD who undergo DBS have a higher incidence of neuropsychiatric complications. Age, operation time, fasting time, H-FRAT score, and UPDRS III score were the independent risk factors. The GBM model incorporating these predictors exhibited excellent discrimination and clinical utility, achieving 80.0% accuracy in preoperative risk stratification.