Predictors of one-year adverse outcomes after laparoscopic resection for hepatocellular carcinoma: Development and validation of an early-warning model

Abstract

BACKGROUND

The global burden of primary liver cancer (PLC) continues to rise. Although minimally invasive, especially laparoscopic, resection is increasingly performed for early-stage disease, 1-year adverse outcomes (recurrence, metastasis, or mortality) remain common. Widely used scores, such as the albumin-bilirubin grade, primarily assess hepatic reserve and may not fully reflect tumor biology or systemic inflammation for individualized early prognostic warning. This study aimed to develop and validate a least absolute shrinkage and selection operator (LASSO)-based model to predict 1-year adverse outcomes after minimally invasive PLC resection.

AIM

To identify predictors of short-term (1-year) adverse outcomes following minimally invasive PLC resection and construct an individualized postoperative prognostic model using LASSO regression.

METHODS

This retrospective study included patients with PLC who underwent minimally invasive resection at The Affiliated Suqian Hospital of Xuzhou Medical University between January 2019 and January 2023. Prognostic predictors were identified using LASSO regression and incorporated into a logistic regression model. Model performance and clinical utility were evaluated using receiver operating characteristic curves, calibration plots, and decision curve analysis. The dataset was randomly divided into training (n = 277) and internal validation (n = 144) cohorts. An external validation cohort of 138 patients with PLC (February 2023 to June 2024) was used to assess generalizability.

RESULTS

Receiver operating characteristic analysis indicated good performance of the logistic regression model based on six predictors, white blood cell count, tumor diameter, vascular invasion, portal vein infiltration, cirrhosis, and alpha-fetoprotein, with area under the curve (AUC) values of 0.756 [95% confidence interval (CI): 0.687-0.824] and 0.750 (95%CI: 0.659-0.841) in the training and internal validation cohorts, respectively. The model exhibited strong calibration (training, P = 0.6951; external validation, P = 0.5223) and clear net clinical benefit across risk thresholds. External validation further supported its generalizability (n = 138; AUC = 0.735, 95%CI: 0.640-0.830). Compared with albumin-bilirubin, the LASSO-based risk score showed higher though non-significant AUCs in the training (0.756 vs 0.691; DeLong P = 0.206) and external (0.735 vs 0.717; P = 0.803) cohorts and comparable performance in the internal validation cohort (0.750 vs 0.753; P = 0.968).

CONCLUSION

LASSO regression was used to identify six independent predictors of adverse 1-year outcomes after minimally invasive PLC resection. The resulting risk score model demonstrates reliable discrimination, calibration, and clinical utility for individualized prognostic assessment.

Key Words: Primary liver cancer; Minimally invasive resection; Outcomes; Predictors; Least absolute shrinkage and selection operator regression; Prognostic model

Core Tip: Preoperatively assess white blood cell count, alpha-fetoprotein, and liver function; evaluate tumor size and vascular or portal vein invasion through imaging; document cirrhosis status. Apply the least absolute shrinkage and selection operator-based model to calculate individualized 1-year risk and guide tailored postoperative follow-up. High-risk patients should undergo multidisciplinary team review and early intervention. Combine risk score output with albumin-bilirubin scoring to assess hepatic reserve, continuously refining thresholds to enhance accuracy, generalizability, and clinical utility. This integrative approach enables early identification of recurrence risk and supports precision management after minimally invasive resection for hepatocellular carcinoma.

INTRODUCTION

Globally, primary liver cancer (PLC) ranks as the sixth most common malignancy and the fourth leading cause of cancer-related mortality[1]. Hepatocellular carcinoma (HCC) and intrahepatic cholangiocarcinoma account for 75%-85% and 10%-15% of cases, respectively, while combined hepatocellular-cholangiocarcinoma remains rare[2]. Men are more susceptible to liver cancer (LC) and show higher mortality rates. The global incidence continues to rise, with over 1 million cases projected by 2025, more than 70% in Asia and approximately 50%in China, posing a major public health challenge[3,4].

LC onset is often insidious, with nonspecific early symptoms, leading most patients to present at intermediate or advanced stages. Early screening of high-risk groups and effective prevention are thus critical to reducing mortality and improving survival[5]. The primary risk factors include hepatitis B virus (HBV) infection, especially prevalent in East Asia and sub-Saharan Africa. In China about 90% of LC patients are infected with HBV, and 85%-90% exhibits cirrhosis[6,7]. Hepatitis C virus infection also contributes substantially to LC development[8].

For early-stage LC, first-line curative options include partial hepatectomy, liver transplantation, and local ablation[9]. Given limited donor availability, radical resection remains preferred. Minimally invasive approaches, laparoscopic or robot-assisted, are increasingly adopted for LC, offering reduced trauma, faster recovery, and fewer complications, thereby improving quality of life and postoperative recovery[10]. However, postoperative recurrence remains high (60%-80%), significantly affecting long-term prognosis. Recurrence is categorized as early (< 2 years) or late (> 2 years), attributed respectively to micrometastasis of the primary lesion or new tumor formation within the altered tumor microenvironment[11]. Therefore, early identification of high-risk patients and tailored management are crucial to prolonging survival.

Accurate prognostic assessment is vital for LC management, as it supports individualized treatment, postoperative optimization, and better outcomes[12]. Logistic and least absolute shrinkage and selection operator (LASSO) regression are commonly used to construct predictive models integrating clinical, laboratory, and imaging parameters. LASSO’s efficient variable selection enhances model accuracy and simplicity, explaining its wide use in complex disease modeling[13]. However, few studies have examined short-term (1-year) prognosis in stage I-II LC after minimally invasive resection. Existing studies are limited by small sample sizes and insufficient external validation, constraining model generalizability.

This study aimed to identify risk factors for poor short-term prognosis in patients with PLC undergoing minimally invasive resection and to develop an individualized LASSO-based warning model for postoperative outcome prediction. Its novelty lies in focusing on minimally invasive surgery, optimizing variable selection through LASSO, ensuring stability and practicability via training-validation comparisons, and constructing a clinically applicable nomogram. Accurate prognostic evaluation enables clinicians to identify high-risk patients early, implementing timely interventions, and adopt aggressive management strategies to reduce recurrence and improve survival.

MATERIALS AND METHODS

Research design

This retrospective study aimed to identify the risk factors for poor 1-year prognosis after minimally invasive resection in patients with PLC and to construct an individualized warning model using LASSO regression.

Data source

Clinical data were obtained from the electronic medical record system of The Affiliated Suqian Hospital of Xuzhou Medical University. The study included patients with PLC who underwent minimally invasive resection between January 2019 and January 2023. An additional 138 patients with PLC treated between February 2023 and June 2024 were retrospectively included as the external validation cohort.

Inclusion and exclusion criteria

Inclusion criteria: (1) Patients met the diagnostic criteria in the Diagnosis and Treatment Specifications for Primary Liver Cancer (2017 Edition)[14], with disease stage I-II; (2) Postoperative pathology confirmed PLC (HCC); (3) All patients underwent laparoscopic surgery; (4) Child-Pugh grade A or B; and (5) Complete clinical and follow-up data.

Exclusion criteria: (1) Received preoperative radiotherapy, chemotherapy, or other antitumor therapy; (2) Tumors invading adjacent organs or exhibiting distant metastasis; (3) Had dysfunction of vital organs (kidneys, heart, or lungs); (4) Had coagulation disorders; (5) Had concurrent malignancies; (6) Died during postoperative hospitalization; or (7) Underwent re-resection for recurrent LC.

Data collection

Baseline clinical data were extracted from the hospital’s electronic medical records, including sex, smoking and alcohol history, hypertension, diabetes, HBV infection status, Child-Pugh classification[15], tumor diameter, tumor number, differentiation degree, vascular invasion, hepatic vein infiltration, portal vein infiltration, and cirrhosis status. Sex was categorized as male or female; smoking, alcohol use, hypertension, and diabetes as yes and no; and HBV infection as infected or uninfected. Child-Pugh class was recorded as grade A or B. Tumor diameter was classified as ≥ 5 cm or < 5 cm, tumor number as single or multiple, and differentiation degree as poor, moderate, or well. Vascular invasion, hepatic vein infiltration, portal vein infiltration, and cirrhosis were dichotomized as present/absent or cirrhosis/non-cirrhosis. Laboratory parameters included white blood cell count (WBC), platelet count (PLT), albumin (ALB), alanine aminotransferase (ALT), aspartate aminotransferase (AST), total bilirubin (TBIL), and alpha-fetoprotein (AFP; ≥ 400 μg/L vs < 400 μg/L, threshold: 400 μg/L). The ALB-bilirubin (ALBI) score was calculated using the standard formula: ALBI = [Log₁₀TBIL (μmol/L) × 0.66] + [ALB (g/L) × (-0.085)].

Criteria for adverse prognosis

Patient were grouped based on 1-year postoperative outcomes. The good prognosis group included those without recurrence, metastasis, or death within one year, whereas the poor-prognosis group included patients who experienced any of these events.

LASSO-logistic regression

In this study, LASSO regression[16] was used for variable selection and regularization in the presence of numerous predictors by introducing an L1 penalty term. This penalty shrinks less significant coefficients toward zero, thereby isolating key prognostic factors. Such an approach reduces model complexity, minimizes overfitting, and enhances generalizability. During model construction, optimal λ values were determined using 10-fold cross-validation.

Statistical analysis

Descriptive statistics summarized patient baseline characteristics; continuous variables were expressed as mean ± SD or median (interquartile range), and categorical variables as n (%). Intergroup differences (good vs poor prognosis) were analyzed using independent-samples t-test or Mann-Whitney U tests for continuous data and χ² or Fisher’s exact tests for categorical data. LASSO regression, through its inherent penalty term, was employed to identify prognosis-related predictors, which were then entered into a logistic regression model to calculate poor-prognosis probabilities. Model performance was evaluated using the area under the receiver operating characteristic (ROC) curve (AUC), calibration curves, and decision curve analysis (DCA) to assess discrimination, accuracy, and clinical net benefit, respectively. All analyses were performed using SPSS version 26.0, and visualization was conducted in R version 4.1.3. A P < 0.05 was considered statistically significant.

RESULTS

Clinical feature differences between prognosis groups

Significant intergroup differences were observed across most clinical variables. The poor-prognosis group included more male, smoking, and alcohol-using patients, whereas the good prognosis group had a higher prevalence of hypertension and diabetes. Poor-prognosis patients more frequently exhibited HBV infection and Child-Pugh grade B liver function. Larger tumor diameters, multiple nodules, and poor differentiation were also significantly associated with poor outcomes. Furthermore, vascular invasion and hepatic or portal vein infiltration were more common in the poor-prognosis group, underscoring their prognostic significance. Cirrhosis and elevated AFP levels were likewise more frequent among poor-prognosis patients. Although WBC and PLT values were slightly higher in the poor-prognosis group, other indicators, ALT, AST, and TBIL differed insignificantly (Figure 1).

Figure 1 Baseline data of patients in the poor and good prognosis groups. In the figure, orange represents the following: Male (sex), history of smoking (smoking history), history of alcohol consumption (alcohol consumption), history of hypertension (history of hypertension), diabetes (diabetes), hepatitis B (hepatitis B), Child-Pugh classification A (Child-Pugh classification), tumor diameter ≥ 5 cm (tumor diameter), single tumor (tumor number), poorly differentiated tumor (degree of differentiation), vascular invasion present (vascular invasion), hepatic vein infiltration present (hepatic vein infiltration), portal vein infiltration present (portal vein infiltration), cirrhosis present (cirrhosis), alpha-fetoprotein (AFP) ≥ 400 μg/L (AFP). Green represents the following: Female (sex), no history of smoking (smoking history), no history of alcohol consumption (alcohol consumption), no history of hypertension (history of hypertension), no diabetes (diabetes), no hepatitis B (hepatitis B), Child-Pugh classification B (Child-Pugh classification), tumor diameter < 5 cm (tumor diameter), multiple tumors (tumor number), moderately to well differentiated tumor (degree of differentiation), no vascular invasion (vascular invasion), no hepatic vein infiltration (hepatic vein infiltration), no portal vein infiltration (portal vein infiltration), no cirrhosis (cirrhosis), AFP < 400 μg/L (AFP). AFP: Alpha-fetoprotein; WBC: White blood cell; PLT: Platelet; ALB: Albumin; ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; TBIL: Total bilirubin.

Correlation of clinical characteristics with laboratory indices

Correlation analyses between clinical and laboratory parameters revealed small correlation coefficients for most variables, indicating no significant linear relationships. For example, WBC and PLT had near-zero correlations with other factors such as hepatic vein infiltration and tumor number. Similarly, ALT and AST showed weak association with liver function-related indicators such as TBIL and Child-Pugh classification. Most coefficients ranged from -0.1 to 0.1, suggesting generally weak correlations. Therefore, no strong linear relationships were detected between clinical characteristics and laboratory indices, indicating that other nonlinear factors may have greater prognostic influence (Figure 2).

Figure 2 Correlation analysis of baseline data. The figure shows the correlation coefficients between each clinical characteristic and laboratory indicator. The color depth represents the correlation coefficient’s magnitude, with orange and blue indicating positive and inverse correlations, respectively. A darker color signifies a stronger association. AFP: Alpha-fetoprotein; WBC: White blood cell; PLT: Platelet; ALB: Albumin; ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; TBIL: Total bilirubin.

Baseline characteristic comparison among training, validation, and external validation groups

When comparing baseline characteristics among the training (n = 277), validation (n = 144), and external validation (n = 138) cohorts, no statistically significant differences were observed across all evaluated variables. Specifically, sex distribution, smoking and alcohol consumption histories, hypertension and diabetes prevalence, HBV infection, Child-Pugh classification, tumor diameter and number, differentiation degree, vascular invasion, hepatic and portal vein infiltration, cirrhosis, and AFP levels showed no statistical differences (all P > 0.05). Similarly, laboratory indices, including WBC count, PLT count, ALB, ALT, AST, and TBIL, were comparable among groups (all P > 0.05). These results confirm that the cohorts had well balanced and statistically comparable baseline characteristics (Table 1).

Table 1 Baseline data comparison (training vs validation groups).

Variable	Training group (n = 277)	Validation group (n = 144)	External validation group (n = 138)	Statistical values	P value
Sex
Male	182	96	91	0.040	0.980
Female	95	48	47
History of smoking
Yes	136	78	67	1.190	0.552
No	141	66	71
History of alcohol consumption
Yes	130	65	62	0.204	0.903
No	147	79	76
History of hypertension
Yes	89	54	47	1.218	0.544
No	188	90	91
History of diabetes
Yes	33	16	16	0.060	0.971
No	244	128	122
Hepatitis B
Yes	193	106	89	2.778	0.249
No	84	38	49
Child-Pugh classification
A	170	92	85	0.273	0.872
B	107	52	53
Tumor diameter
≥ 5 cm	120	55	54	1.285	0.526
< 5 cm	157	89	84
Tumor number
Single	221	111	112	0.759	0.684
Multiple	56	33	26
Degree of differentiation
Poorly differentiated	36	20	20	0.190	0.909
Moderately or well differentiated	241	124	118
Vascular invasion
Yes	66	30	35	0.853	0.653
No	211	114	103
Hepatic vein infiltration
Yes	18	11	8	0.400	0.819
No	259	133	130
Portal vein infiltration
Yes	37	20	18	0.045	0.978
No	240	124	120
Cirrhosis
Yes	165	80	79	0.664	0.717
No	112	64	59
AFP
≥ 400 μg/L	75	45	37	0.965	0.617
< 400 μg/L	202	99	101
WBC (× 109/L)	6.62 ± 1.28	6.74 ± 1.24	6.55 ± 1.33	0.844	0.431
PLT (× 109/L)	137.67 ± 24.72	138.74 ± 23.50	138.56 ± 22.91	0.119	0.888
ALB (g/L)	39.03 ± 7.66	39.74 ± 7.73	38.90 ± 7.83	0.520	0.595
ALT (U/L)	95.23 ± 17.37	95.34 ± 16.35	92.87 ± 17.04	1.037	0.355
AST (U/L)	100.87 ± 18.07	100.05 ± 17.43	100.62 ± 17.58	0.100	0.905
TBIL (μmol/L)	19.42 (16.69, 21.34)	19.41 (16.59, 22.29)	19.16 (15.97, 21.18)	0.873	0.646

WBC: White blood cell; AFP: Alpha-fetoprotein; PLT: Platelet count; ALB: Albumin; ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; TBIL: Total bilirubin.

Baseline characteristic comparison between prognosis groups in the training cohort

In the training cohort, significant intergroup differences were observed between the good (n = 210) and poor prognosis (n = 67) groups. Tumor diameter, vascular invasion, portal vein infiltration, cirrhosis, and AFP levels differed significantly (P < 0.05). Patients with tumor diameter ≥ 5 cm, vascular invasion, portal vein infiltration, cirrhosis, or AFP ≥ 400 μg/L were more frequent in the poor-prognosis group. WBC count was also statistically higher among poor-prognosis patients (P = 0.011). No significant differences were found in sex, smoking or alcohol history, hypertension, diabetes, HBV infection, Child-Pugh classification, tumor number, differentiation degree, hepatic vein infiltration, PLT, ALB, ALT, AST, or TBIL (P > 0.05; Table 2).

Table 2 Baseline data comparison (good vs poor prognosis groups in the training group).

Variable	Poor prognosis group (n = 67)	Good prognosis group (n = 210)	Statistical values	P value
Sex
Male	46	136	0.342	0.559
Female	21	74
History of smoking
Yes	30	106	0.660	0.416
No	37	104
History of alcohol consumption
Yes	32	98	0.024	0.876
No	35	112
History of hypertension
Yes	21	68	0.025	0.874
No	46	142
History of diabetes
Yes	10	23	0.764	0.382
No	57	187
Hepatitis B
Yes	46	147	0.043	0.835
No	21	63
Child-Pugh classification
A	41	129	0.001	0.973
B	26	81
Tumor diameter
≥ 5 cm	37	83	5.099	0.024
< 5 cm	30	127
Tumor number
Single	48	173	3.632	0.057
Multiple	19	37
Degree of differentiation
Poorly differentiated	13	23	3.208	0.073
Moderately or well differentiated	54	187
Vascular invasion
Yes	23	43	5.370	0.020
No	44	167
Hepatic vein infiltration
Yes	6	12	0.878	0.349
No	61	198
Portal vein infiltration
Yes	15	22	6.228	0.013
No	52	188
Cirrhosis
Yes	47	118	4.109	0.043
No	20	92
AFP
≥ 400 μg/L	33	42	22.015	< 0.001
< 400 μg/L	34	168
WBC (× 109/L)	6.96 ± 1.25	6.51 ± 1.27	2.583	0.011
PLT (× 109/L)	141.88 ± 25.84	136.32 ± 24.27	1.555	0.123
ALB (g/L)	38.28 ± 8.21	39.27 ± 7.48	-0.886	0.378
ALT (U/L)	96.51 ± 17.76	94.83 ± 17.27	0.679	0.499
AST (U/L)	100.42 ± 18.82	101.01 ± 17.87	-0.224	0.823
TBIL (μmol/L)	19.39 ± 3.73	19.22 ± 3.61	0.33	0.742

AFP: Alpha-fetoprotein; WBC: White blood cell count; PLT: Platelet count; ALB: Albumin; ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; TBIL: Total bilirubin.

Feature selection and prediction model construction using LASSO regression

Using LASSO regression, six prognostic variables were identified: WBC count, tumor diameter, vascular invasion, portal vein infiltration, cirrhosis, and AFP level (Figure 3). The optimal penalty parameter was λ = 0.0035191, determined through cross-validation. Regression coefficients were 0.216 for WBC count, 0.644 for tumor diameter, 0.813 for vascular invasion, 0.883 for portal vein infiltration, 0.431 for cirrhosis, and 1.385 for AFP level. The model formula was Logit(P) = -3.998 + 0.216 × WBC + 0.644 × tumor diameter + 0.813 × vascular invasion + 0.883 × portal vein infiltration + 0.431 × cirrhosis + 1.385 × AFP, where P represents the probability of poor prognosis. This model enables individualized prognosis prediction based on the identified variables.

Figure 3 Least absolute shrinkage and selection operator regression screening of risk characteristic variables for poor prognosis. A: Least absolute shrinkage and selection operator regression path diagram. As Log(λ) augments, the regression coefficients of the characteristic variables gradually contract to zero. When the λ value is small, most variables are included in the model. As the λ value escalates, the coefficients of some variables incline towards zero, and ultimately 6 key characteristic variables are screened out; B: Cross-validation curve diagram. The vertical axis represents the binomial deviation, and the horizontal axis denotes Log(λ). The optimal λ value is selected via 10-fold cross-validation. The dotted line position corresponds to the minimum value of λ (λ = 0.0035191), and the model corresponding to this value has the lowest deviation. Six characteristic variables are screened out at this point for the final model construction.

ROC, DCA, and calibration curve validation of the LASSO model in the training cohort

In the training cohort, ROC curve (Figure 4A) showed an AUC of 0.756 [95% confidence interval (CI): 0.687-0.824; P = 1.89 × 10^-13], indicating good discriminatory power in distinguishing good from poor-prognosis patients. DCA curve (Figure 4B) demonstrated a high clinical net benefit across risk thresholds of 0%-84%, with a maximum benefit of 24.18%. The goodness-of-fit test of the calibration curve (Figure 4C) showed that the model was well calibrated (χ² = 5.5714, P = 0.6951).

Figure 4 Receiver operating characteristic curve, decision curve, and calibration curve validation of the least absolute shrinkage and selection operator model (training group). A: The receiver operating characteristic curve demonstrates the model’s good discriminatory power (area under the curve = 0.756) in the training group; B: Decision curve analysis evaluation of the model’s clinical net benefit showed a maximum net benefit of 24.18% within the 0%-84% threshold range; C: The calibration curve shows a good fit and calibration between model predictions and the actual occurrence rate. The χ² value of the goodness of fit test is 5.5714, and the P value is 0.6951. TPR: True positive rate; FPR: False positive rate; AUC: Area under the curve; CI: Confidence interval.

ROC, DCA, and calibration curve validation of the LASSO model in the validation cohort

In the validation cohort, the LASSO model was evaluated and validated using ROC curves, DCA, and calibration plots. The ROC curve (Figure 5A), yielded an AUC of 0.750 (95%CI: 0.659-0.841; P = 4.76E-8), confirming strong discrimination between favorable and unfavorable clinical outcomes. The DCA curve (Figure 5B) showed robust net benefit across risk thresholds of 0%-96%, while the goodness-of-fit test result of the calibration plot (Figure 5C) indicated strong model calibration (χ² = 8.2821, P = 0.4064).

Figure 5 Receiver operating characteristic, decision curve analysis, and calibration curve validation of the least absolute shrinkage and selection operator model (validation group). A: The receiver operating characteristic curve demonstrates the model’s good discrimination ability in the validation group (area under the curve = 0.750); B: Decision curve analysis reveals the model’s significant clinical net benefit within the 0%-96% threshold range; C: The calibration curve shows a good fit and favorable calibration between model predictions and the actual occurrence rate. The χ² value of the goodness-of-fit test is 8.2821, and the P value is 0.4064. TPR: True positive rate; FPR: False positive rate; AUC: Area under the curve; CI: Confidence interval.

ROC, DCA, and calibration curve validation of the LASSO model in the external validation cohort

In the external validation cohort, the LASSO model achieved an AUC of 0.735 (95%CI: 0.640-0.830, P = 9.58E-07) in ROC analysis (Figure 6A), demonstrating reliable discrimination between patients at high and low risk of poor outcomes. The DCA curve (Figure 6B) revealed a positive clinical net benefit within the 0%-90% threshold probability range, peaking at 77.53%. The calibration plot (Figure 6C) showed strong agreement between predicted and observed outcomes, supported by goodness-of-fit testing (χ² = 7.1336, P = 0.5223).

Figure 6 Receiver operating characteristic, decision curve analysis, and calibration curve validation of the least absolute shrinkage and selection operator model (external validation group). A: The receiver operating characteristic curve shows an area under the curve of 0.735 (95%confidence interval: 0.640-0.830) of the least absolute shrinkage and selection operator model in the external validation group, indicating good discriminatory ability; B: The decision curve analysis curve evaluates the model’s clinical net benefit, revealing a significant net benefit within the 0%-90% threshold probability range, with a maximum net benefit of 77.53%; C: The calibration curve illustrates the agreement between the predicted probabilities and actual outcomes. Good model calibration is evidenced by a non-significant goodness-of-fit result (χ² = 7.1336, P = 0.5223). TPR: True positive rate; FPR: False positive rate; AUC: Area under the curve; CI: Confidence interval.

Comparison of ROC curve parameters of the LASSO risk score between training and validation cohorts

The ROC curve parameters of the LASSO-derived risk score were consistent across the training and validation cohorts, confirming stable discriminatory performance. The AUC values were 0.756 (95%CI: 0.687-0.824) for the training cohort and 0.750 (95%CI: 0.659-0.841) for the validation cohort. In the training cohort, sensitivity, specificity, and Youden index were 74.63%, 67.62%, and 42.25%, respectively; in the validation cohort, these values were 84.85%, 54.95%, and 39.80%. The optimal cut-off values were -1.256 (training) and -1.604 (validation). Model accuracy, precision, and F1 score in the training cohort were 69.31%, 74.63%, and 54.05%, respectively, compared with 61.81%, 84.85%, and 50.45% in the validation cohort (Table 3). These findings underscore the model’s consistent and robust discrimination across datasets.

Table 3 Receiver operating characteristic curve parameters of the training and validation groups.

Marker	AUC	95%CI	Specificity, %	Sensitivity, %	Youden index, %	Cut off	Accuracy, %	Precision	F1 score, %
LASSO risk score (training group)	0.756	0.687-0.824	67.62	74.63	42.25	-1.256	69.31	74.63	54.05
LASSO risk score (validation group)	0.75	0.659-0.841	54.95	84.85	39.80	-1.604	61.81	84.85	50.45

LASSO: Least absolute shrinkage and selection operator; AUC: Area under the curve; CI: Confidence interval.

Incremental discrimination over ALBI across training, validation, and external cohorts

Compared with the ALBI score, the LASSO-based risk score consistently demonstrated higher discriminatory ability across all cohorts. In the training cohort, risk score achieved a higher AUC than ALBI (0.756 vs 0.691; DeLong P = 0.206, Figure 7A). The improvement persisted in the validation (0.750 vs 0.753; DeLong P = 0.968, Figure 7B) and external (0.735 vs 0.717; DeLong P = 0.803, Figure 7C) cohorts. Although statistical significance was reached only in the training cohort, the direction and magnitude of improvement were consistent, supporting the LASSO model’s incremental predictive value beyond ALBI.

Figure 7 Comparison of the least absolute shrinkage and selection operator-based risk score vs albumin bilirubin in discriminating 1-year adverse outcomes across three cohorts. A: Training cohort: Risk score (RISK) outperformed albumin bilirubin (ALBI); B: Validation cohort: RISK outperformed ALBI; C: External cohort: RISK outperformed ALBI. RISK: Risk score; ALBI: Albumin bilirubin; TPR: True positive rate; FPR: False positive rate; AUC: Area under the curve; CI: Confidence interval.

Construction of the nomogram model and assessment of patient prognostic risk

A prognostic nomogram model (Figure 8) was developed using the six LASSO-identified variables, WBC count, tumor diameter, vascular invasion, portal vein infiltration, cirrhosis, and AFP, to estimate individualized prognosis. Each variable contributed a weighted score, and the total score was used to calculate the probability of poor prognosis. For instance, of the two patients randomly selected, patient 1 had a total score of 418, corresponding to a 100% poor prognosis probability, whereas patient 2 scored 268 points, corresponding to a 32.5% poor prognosis (Figure 9; Table 4). This model effectively distinguished patients with different prognoses, demonstrating high predictive accuracy.

Figure 8 Construction of a nomogram model based on the variables screened by least absolute shrinkage and selection operator regression. WBC: White blood cell; AFP: Alpha-fetoprotein.

Figure 9 Calculation of nomogram scores for randomly selected patients with poor and good prognoses. A: Nomogram score chart of patient 1, with a total score of 418 and a 100% probability of poor prognosis; B: Nomogram score chart of patient 2 with a total score of 268 and a poor prognosis of 32.5%. WBC: White blood cell; AFP: Alpha-fetoprotein.

Table 4 Calculation of patient score.

Variable	Patient 1	Patient 2
WBC (× 109/L)	9.52	5.5
Tumor diameter	< 5 cm	≥ 5 cm
Vascular invasion	Yes	No
Portal vein infiltration	Yes	No
Cirrhosis	Yes	Yes
AFP	≥ 400 μg/L	< 400 μg/L
Total score	418	268
Probability of poor prognosis	100%	32.5%

WBC: White blood cell; AFP: Alpha-fetoprotein.

DISCUSSION

Through retrospective analysis, six risk factors, WBC count, tumor diameter, vascular invasion, portal vein infiltration, liver cirrhosis, and AFP level, were identified as strongly associated with poor prognosis after minimally invasive resection for PLC. The logistic regression model constructed based on these variables achieved AUCs of 0.756 and 0.750 in the training and validation cohorts, respectively, demonstrating strong discriminatory ability in predicting one-year postoperative outcomes. The DCA and calibration curves further confirmed the model’s clinical applicability and good calibration, supporting its potential for use in clinical decision-making.

The six prognostic factors identified in this study align closely with established findings. For instance, AFP, a widely recognized HCC biomarker, correlates with tumor aggressiveness and recurrence. Li et al[17] developed a prognostic model for elderly patients with PLC using the SEER database and identified AFP as a key prognostic indicator. Similarly, tumor diameter and vascular invasion have been repeatedly validated as major prognostic determinants in LC. Nevola et al[18] observed higher early recurrence and metastasis rates in patients with larger tumors, whereas vascular invasion remains a hallmark of tumor invasion.

This study specifically focused on patients undergoing minimally invasive resection, a population rarely examined in prior research, offering new insights into prognostic evaluation following this approach. While previous models have been developed for non-cirrhotic HCC[19] or LC with microvascular invasion[20], this study employed LASSO regression to manage multivariate complexity and enhance model generalizability. Compared with traditional logistic regression model, LASSO mitigates overfitting and improves predictive accuracy.

Elevated WBC count, an inflammation marker, may reflect an exacerbated tumor-related inflammatory state, which has been linked to poor outcomes across malignancies. Zhan et al[21] emphasized the prognostic role of inflammatory markers in LC. Larger tumor diameter typically correlates with higher recurrence and metastasis risks due to increased tumor burden and invasiveness. Brustia et al[22] identified tumor diameter as a key predictor of postoperative complications and long-term prognosis. Vascular invasion, indicative of high metastatic potential, consistently predicts unfavorable outcomes. Portal vein infiltration further suggests advanced tumor invasiveness, contributing to greater recurrence risk and poor survival, as confirmed by Yang et al[23]. Liver cirrhosis, reflecting impaired hepatic function, compromises postoperative recovery and long-term outcomes and is therefore an essential background factor in LC progression. Prior studies have associated liver cirrhosis with delayed recovery and increased recurrence. Elevated AFP serves not only as a diagnostic biomarker but also an indicator of tumor invasiveness and recurrence potential, mirroring the biological behavior of malignant hepatocytes[24,25]. Liu et al[26] also confirmed the importance of AFP in LC. Even among AFP-negative cases, factors such as tumor diameter and vascular invasion remain critical predictors. In the logistic regression model, AFP exhibited the strongest association with poor prognosis, followed by portal vein infiltration and vascular invasion, highlighting their dominant predictive roles. WBC count and cirrhosis had smaller but significant effects, underscoring their value in comprehensive prognostic assessment.

Thus, the LASSO regression-based early-warning model can effectively identify high-risk patients soon after surgery and assist clinicians in developing individualized follow-up and management plans. For instance, high-risk patients may require more frequent imaging and proactive adjuvant therapy to lower recurrence or metastasis risk[27]. Early identification and intervention can significantly reduce postoperative recurrence, prolong survival, and enhance well-being[28,29], optimizing treatment outcomes while easing healthcare system burdens. The SEER-based prediction model has also proven useful for clinical decision-making, improving treatment precision and individualization[21]. Similar models by He et al[30] and Zhou et al[31] achieved good predictive ability; however, our model demonstrated higher AUC and C-index values, particularly in the validation cohort, reflecting superior generalization.

Compared with existing models, our individualized early-warning model offers better accuracy and clinical applicability. Li et al[17] developed a SEER-based prognostic model for elderly patients with PLC with C-indices of 0.747 and 0.773 in training and validation cohorts, respectively, while our model achieved comparable AUC values (0.756 and 0.750). Zhou et al[31] established a LASSO Cox regression model for AFP-negative patients with LC (C-index = 0.752 in the validation cohort), also comparable to our model results. The cancer-specific survival model by Yang et al[23] for LC patients with microvascular invasion yielded C-indices of 0.785 and 0.776, indicating slightly higher predictive accuracy. Nevertheless, our external validation demonstrated an AUC of 0.735 (95%CI: 0.640-0.830) for the LASSO model, confirming consistent and robust discrimination in an independent PLC cohort. The model’s net clinical benefit was 0%-90% threshold probabilities, peaking at 77.53%. Its calibration curve also indicated close agreement between predicted and observed outcomes (χ² = 7.1336, P = 0.5223), confirming good calibration. Collectively, these results shows that our model reliably predicts postoperative prognosis and remains competitive with comparable models. Its strong applicability across different populations and treatment modalities further supports its clinical utility.

This study offers several strengths: A large sample size (277 patients in the training and 144 patients in the validation cohorts) that enhances statistical power and result reliability; variable screening through LASSO regression, which reduces model complexity and boosts generalizability; and model validation using ROC, DCA, and calibration analyses to ensure predictive accuracy and clinical value. However, certain limitations exist. First, all data were sourced from a single hospital, which may restrict the model’s generalizability to other regions or healthcare settings. Second, as a retrospective study, selection bias may affect result authenticity and reliability. Third, the model lacks validation on other independent datasets, necessitating further testing in diverse populations. Finally, genetic background and lifestyle factors were not included, possibly omitting relevant prognostic factors. As Li et al[17] noted, multicenter and prospective studies are essential to enhance the model’s universality and precision.

Future research should focus on external validation across hospitals and regions to assess model performance in diverse populations and enhance its universality and reliability. Prospective cohort studies are also warranted to minimize biases inherent in retrospective designs and to further verify predictive accuracy. Additionally, refining the model by incorporating more clinical and molecular biology indicators, such as gene expression profiles and tumor microenvironmental features, could improve precision. Translating the model into a clinical decision support system and evaluating its utility and impact in real-world clinical settings would facilitate broader clinical adoption. Furthermore, exploring the biological mechanisms underlying the identified LC prognostic factors could clarify their roles in tumor progression and provide a theoretical basis for targeted therapy. Yang et al[23] and Zhan et al[21] similarly emphasized integrating clinical factors with multiple biomarkers to enhance model comprehensiveness and prognostic accuracy.

CONCLUSION

This study identified six risk factors significantly associated with poor prognosis in PLC after minimally invasive resection using LASSO regression and developed an individualized early-warning model with strong predictive performance and calibration. The model demonstrated high discriminatory ability and clinical utility in both training and validation cohorts and may aid in early identification of high-risk patients, optimizing postoperative management and follow-up strategies to extend survival and improve quality of life. Future multicenter and prospective studies are needed to further validate and optimize this model, ensuring its broad applicability and reliability across diverse populations.