Machine learning model using immune indicators to predict outcomes in early liver cancer

Abstract

BACKGROUND

Patients with early-stage hepatocellular carcinoma (HCC) generally have good survival rates following surgical resection. However, a subset of these patients experience recurrence within five years post-surgery.

AIM

To develop predictive models utilizing machine learning (ML) methods to detect early-stage patients at a high risk of mortality.

METHODS

Eight hundred and eight patients with HCC at Beijing Ditan Hospital were randomly allocated to training and validation cohorts in a 2:1 ratio. Prognostic models were generated using random survival forests and artificial neural networks (ANNs). These ML models were compared with other classic HCC scoring systems. A decision-tree model was established to validate the contribution of immune-inflammatory indicators to the long-term outlook of patients with early-stage HCC.

RESULTS

Immune-inflammatory markers, albumin-bilirubin scores, alpha-fetoprotein, tumor size, and International Normalized Ratio were closely associated with the 5-year survival rates. Among various predictive models, the ANN model generated using these indicators through ML algorithms exhibited superior performance, with a 5-year area under the curve (AUC) of 0.85 (95%CI: 0.82-0.88). In the validation cohort, the 5-year AUC was 0.82 (95%CI: 0.74-0.85). According to the ANN model, patients were classified into high-risk and low-risk groups, with an overall survival hazard ratio of 7.98 (95%CI: 5.85-10.93, P < 0.0001) between the two cohorts.

CONCLUSION

A non-invasive, cost-effective ML-based model was developed to assist clinicians in identifying high-risk early-stage HCC patients with poor postoperative prognosis following surgical resection.

Key Words: Hepatocellular carcinoma; Inflammation; Machine learning; Prognosis; Artificial neural networks; Immune biomarkers

Core Tip: This study developed a predictive model using machine learning algorithms that integrates immune-inflammatory biomarkers to forecast the long-term prognosis of patients following surgical resection for early-stage hepatocellular carcinoma. This model aims to optimize screening and treatment strategies.

INTRODUCTION

Hepatocellular carcinoma (HCC) is one of the six most prevalent cancers[1] and the third leading cause of cancer-related mortality[2]. China has some of the highest incidence and mortality rates for liver cancer, accounting for half of global cases[3,4]. The Barcelona Clinic Liver Cancer (BCLC) Staging System is the most widely used framework for diagnosing and treating HCC[5]. The optimal candidates for surgical treatment are those with early-stage HCC, classified as BCLC stage 0 or A. Patients with early-stage liver cancer typically have a better prognosis after surgical resection, achieving a 5-year survival rate of 60%-70%[6]. However, the high postoperative recurrence rates of HCC remain a major obstacle to long-term efficacy. To improve the prognosis of patients with early-stage HCC, it is necessary to develop models that can identify those with poor prognoses, enabling stratified and personalized treatment and follow-up strategies.

Chronic inflammation is linked to the development and advancement of tumors[7]. Recently, peripheral blood immune indicators, such as neutrophil-to-lymphocyte ratio (NLR), platelet-to-lymphocyte ratio (PLR), and lymphocyte-to-monocyte ratio (LMR), have garnered extensive attention and have been used to predict survival in various tumors and inflammation-related diseases[8-10]. However, the relationship between these combinations of immune markers and the outcomes in patients with early-stage HCC require further investigation.

Machine learning (ML) algorithms are capable of handling large and complex datasets, generating more accurate and personalized predictions through unique training algorithms that better manage nonlinear statistical relationships than traditional analytical methods. Commonly used ML models include artificial neural networks (ANNs) and random survival forests (RSFs), which have shown satisfactory accuracy in prognostic predictions across various cancers and other diseases[11-13]. ANNs have performed well in identifying the progression from liver cirrhosis to HCC and predicting overall survival (OS) in patients with HCC[14,15]. However, no studies have confirmed the ability of ML models to predict post-surgical survival in patients with early-stage HCC. Through ML, a better understanding of the risk factors for early-stage HCC prognosis can be achieved. This aids in surgical decision-making, identifying patients at a high risk of mortality, and selecting subsequent treatment strategies.

In this study, we aimed to establish a 5-year prognostic model for patients with early-stage HCC after surgical resection, based on ML and systemic immune-inflammatory indicators. This model seeks to improve the early monitoring of high-risk patients and provide personalized treatment plans.

MATERIALS AND METHODS

Patients and follow-up

Data for this study were obtained from patients with HCC treated at the Beijing Ditan Hospital, affiliated with the Capital Medical University. Initially, we screened patients diagnosed with HCC at BCLC stage 0-A between September 2008 and September 2018. The exclusion criteria for patients included: (1) Other malignancies; (2) BCLC stage B-D; (3) Incomplete data or lack of follow-up; (4) Age < 18 years or > 80 years; (5) Treatments other than surgical resection, such as radiofrequency ablation, trans arterial chemoembolization, or other therapies; and (6) Diagnosis of autoimmune liver disease or acquired immune deficiency syndrome. Based on these criteria, 808 patients were selected for further evaluation and randomly assigned to training and validation groups in a 2:1 ratio. The primary outcome was mortality rate within five years or at the conclusion of the 5-year follow-up period. Follow-ups were conducted primarily through hospital records and telephone interviews.

This study was approved by the Ethics Committee of Beijing Ditan Hospital, and all patients provided written informed consent. This research was conducted in compliance with the Declaration of Helsinki, as revised in 2008.

BCLC staging and diagnosis

The BCLC staging system[5] incorporates factors related to the tumor stage, liver function, overall physical condition, and cancer-related symptoms. The definition of BCLC stage 0 is very early-stage HCC characterized by a solitary liver tumor with a diameter of ≤ 2 cm, good liver function reserve, no tumor-related symptoms, no vascular invasion, and no extrahepatic metastasis. Early-stage HCC (BCLC stage A) includes solitary or multifocal liver cancer (up to three nodules, each measuring ≤ 3 cm) without major vascular invasion, extrahepatic spread, or cancer-related symptoms, irrespective of tumor size. Intermediate stage HCC (BCLC stage B) is characterized by large or multifocal tumors with vascular invasion or extrahepatic spread. Advanced stage HCC (BCLC stage C) includes patients who present with cancer-related symptoms and/or exhibit vascular invasion or extrahepatic spread. Terminal stage HCC (BCLC stage D) includes patients with widespread tumor invasion and severely impaired liver function. The diagnosis of HCC was based on imaging (hepatic arteriography, magnetic resonance imaging, or liver ultrasound imaging) or histopathological evaluation, which was consistent with the American Association for the Study of Liver Diseases diagnostic guidelines[16]. In this study, we enrolled patients with early-stage HCC (BCLC stage 0 to stage A).

Score calculations

NLR was determined as the ratio of neutrophils to lymphocytes, PLR as the ratio of platelets to lymphocytes, and LMR as the ratio of lymphocytes to monocytes.

Albumin-bilirubin (ALBI) grade was calculated using the following formula[17]: ALBI = (log₁₀ (total bilirubin, μmol/L) × 0.66) + (albumin, g/L × -0.085)

The ALBI grades were classified as follows: (1) Grade I: ≤ -2.60; (2) Grade II: > -2.60 to ≤ -1.39; and (3) Grade III: > -1.39.

RSF model construction

The RSF model was built by performing bootstrapped resampling of the dataset to generate multiple subsets. A total of 500 decision trees were constructed [number of trees (n-tree) = 500], with a minimum node size of 1 (node size = 1) to control the tree depth. At each node split, five candidate variables were randomly selected (mtry = 5). The importance of each variable was assessed based on the increase in out-of-bag (OOB) prediction error. Model performance was evaluated using OOB data to prevent overfitting. Additionally, to minimize the potential impact of feature interactions, we excluded variables likely to exhibit strong interactions, such as NLR with lymphocyte count, neutrophil count, and white blood cell count. This approach ensured the independence and appropriateness of feature selection for the model input.

ANN model building process

The constructed ANN model consisted of layers with five neurons, using the rectified linear unit as the activation function. The Adam optimizer was employed to update the model’s weights, with an initial learning rate set at 0.01. Full-batch training was used, and the training iterations were controlled using the stepmax parameter to ensure that the model stopped before convergence. To prevent overfitting, dropout regularization was applied. Model performance evaluation and hyperparameter tuning were conducted using a five-fold cross-validation. Once the optimal combination of hyperparameters was identified, the model was retrained on the entire training set and then evaluated on the validation cohorts to assess its generalization capability using unseen data.

Statistical analysis

Normally distributed data are reported as mean ± SD and were analyzed using the t-test. For data that were not normally distributed, results are shown as median (interquartile range 25-75) and were analyzed using the Mann-Whitney U test. Categorical variables were examined using the χ² test or Fisher’s exact test. For variables with less than 5% missing values, multiple imputation was applied for data completion. To identify risk factors influencing early HCC patient survival, we employed a RSF model, created using the 'randomForest' package in R. This nonlinear ensemble learning method predicts overall mortality by calculating the cumulative hazard function for each sample and aggregating these based on survival times[18,19]. The ANN model was constructed using the Python package 'PyTorch' and optimized by adding three hidden layers through multiple tests to achieve the best performance. Decision trees were built using the R packages 'rpart' and 'rpart.plot'. To enhance the interpretability of the ANN model, we utilized the SHAP library in Python to assess the contributions of each feature.

The predictive value of each model was assessed using receiver operating characteristic (ROC) curves and the area under the curve (AUC), with comparisons made using the DeLong test[20]. Model calibration was evaluated using calibration plots, and the clinical net benefits of the models were compared using decision curve analysis (DCA). The optimal cutoff point on the ROC curve was determined using the Youden index, which was subsequently used to stratify patients into high-risk and low-risk groups. The 1-, 3-, and 5-year OS rates were calculated using the Kaplan-Meier method and compared using the log-rank test. Statistical analyses were conducted using R (version 4.4.0) and Python (version 3.12.3), with the significance level set at P < 0.05.

RESULTS

Patient demographics and laboratory characteristics

The average age of the patients was 56 years, comprising 606 males and 206 females. Regarding etiology, 630 patients (77.97%) had hepatitis B virus (HBV) infection, 68 (8.41%) had hepatitis C virus infection, 71 (8.78%) had alcoholic hepatitis, and 39 (4.82%) had liver disease due to autoimmune hepatitis, nonalcoholic steatohepatitis, or unknown causes. In terms of liver function, 369 patients (45.66%) were classified as ALBI grade I, with 387 (47.89%) and 52 (6.43%) patients classified as ALBI grades II and III, respectively. A total of 656 patients (81.18%) had liver function classified as Child-Pugh A, and 152 patients (18.81%) as Child-Pugh B. There were no notable differences in the demographic characteristics, laboratory indicators, or treatment methods between the training and validation cohorts (Table 1).

Table 1 Baseline demographics and clinical characteristics of patients with hepatocellular carcinoma in the training and validation cohorts, n (%).

Variables	Total (n = 808)	Training group (n = 539)	Validation group (n = 269)	P values
Patients background
Gender, male/female	602/206	399/140	203/66	0.721
Age, year (Q1, Q3)	56 (50, 62)	56 (50, 62)	56 (50, 63)	0.575
Cirrhosis	665 (82.20)	438 (81.26)	227 (84.38)	0.318
Decompensation	445 (55.07)	286 (53.09)	159 (59.10)	0.12
Smoking	340 (42.07)	215 (39.88)	125 (46.46)	0.087
Drinking	334 (41.33)	221 (41.00)	113 (42.00)	0.843
Family history of HCC	80 (9.90)	57 (10.57)	23 (8.55)	0.433
Diabetes	176 (21.78)	109 (20.22)	67 (24.90)	0.153
Hypertension	223 (27.59)	145 (26.90)	78 (28.99)	0.586
Antiviral	477 (59.03)	315 (58.44)	162 (60.22)	0.682
Ascites				0.106
Mild ascites	181 (22.40)	111 (20.59)	70 (26.02)
Severe ascites	51 (6.31)	40 (7.42)	11 (4.08)
HE	13 (1.61)	7 (1.29)	6 (2.23)	0.377
UGIB	32 (3.96)	21 (3.89)	11 (4.08)	0.856
Laboratory parameters
ALT, U/L, (Q1, Q3)	27.8 (19, 41.95)	28 (19.1, 42.5)	27.2 (18.9, 41)	0.529
AST, U/L, (Q1, Q3)	30.1 (22.8, 44.15)	30.1 (23.1, 44.2)	30.1 (22.5, 43.3)	0.95
TBIL, μmol/L, (Q1, Q3)	15.6 (10.8, 23)	15.9 (11.1, 22.95)	13.9 (10.3, 23.4)	0.114
PLR, (Q1, Q3)	80.5 (59.68, 111.43)	82.02 (59.65, 112.63)	79.01 (59.89, 107.34)	0.352
NLR, (Q1, Q3)	1.92 (1.4, 2.84)	1.91 (1.4, 2.78)	1.92 (1.4, 2.95)	0.548
LMR, (Q1, Q3)	3.63 (2.59, 4.83)	3.61 (2.59, 4.72)	3.67 (2.59, 5)	0.48
ALBI
I	369 (45.66)	248 (46.01)	121 (45.12)	0.866
II	387 (47.89)	258 (47.86)	129 (46.95)
III	52 (6.43)	33 (6.12)	19 (7.06)
Child-Pugh				0.233
Grade A	656 (81.18)	445 (82.56)	211 (78.43)
Grade B	152 (18.81)	94 (17.43)	58 (21.56)
CRP, mg/L	4.2 (1.5, 12.55)	4.2 (1.5, 12.1)	4.5 (1.5, 13.28)	0.665
ALB, g/L, (Q1, Q3)	38.35 (33.27, 42.3)	38.7 (33.45, 42.4)	37.9 (33.1, 42.1)	0.278
GLB, g/L, (Q1, Q3)	30.2 (26.8, 34.23)	30.1 (26.85, 34.1)	30.3 (26.5, 34.8)	0.929
LDH, U/L, (Q1, Q3)	173.35 (149.88, 201.77)	173.5 (149.8, 203.3)	173.2 (150, 196.6)	0.59
Cr, μmol/L, (Q1, Q3)	0.76 (0.66, 0.87)	0.76 (0.64, 0.87)	0.77 (0.67, 0.9)	0.076
HBV-DNA ≥ 250, IU/mL	287 (35.51)	205 (38.03)	82 (30.48)	0.05
Tumor-related indicators
AFP ≥ 400, ng/mL	133 (16.36)	90 (16.69)	43 (15.98)	0.875
Tumor size, ≥ 2 cm	340 (42.07)	232 (43.04)	108 (40.14)	0.478
Tumor multiplicity				0.96
Solidary	726 (89.85)	485 (89.98)	241 (89.59)
Multiple	82 (10.14)	54 (10.02)	28 (10.41)
BCLC				0.546
BCLC 0	43 (5.32)	31 (5.75)	12 (4.46)
BCLC A	765 (94.67)	508 (94.25)	257 (95.54)
Etiology
HBV-infection	630 (77.97)	420 (77.92)	210 (78.06)	0.965
HCV-infection	68 (8.41)	47 (8.71)	21 (7.80)	0.759
AH	71 (8.78)	43 (7.97)	28 (10.40)	0.308
Other	39 (4.82)	29 (5.38)	10 (3.71)	0.387

Data are presented as n (%), or median (interquartile range). HCC: Hepatocellular carcinoma; HE: Hepatic encephalopathy; UGIB: Upper Gastrointestinal Bleeding; ALT: Alanine aminotransferase; AST: Aspartate aminotransferase; TBIL: Total bilirubin; PLR: Platelet-to-lymphocyte ratio; NLR: Neutrophil-to-lymphocyte ratio; LMR: Lymphocyte-to-monocyte ratio; ALBI: Albumin-bilirubin; CRP: C-reactive protein; ALB: Albumin; GLB: Globulin; LDH: Lactate dehydrogenase; Cr: Creatinine; HBV: Hepatitis B virus; AFP: Alpha-Fetoprotein; BCLC: Barcelona Clinic Liver Cancer Staging; HCV: Hepatitis C virus; AH: Alcoholic hepatitis.

RSF combined with ANN modeling

An RSF model was used to identify prognostic factors affecting the 5-year OS of patients with early-stage HCC. Variable importance was utilized to prioritize the variables (Figure 1A). Parameter tuning showed that the error rate of the model stabilized when the n-tree reached 400 (Figure 1B). The most important variables included ALBI grade, two immune markers (PLR and LMR), alpha-fetoprotein (AFP) level, tumor size, and the International Normalized Ratio (Figure 1C). The important variables identified by RSF were subsequently used for ANN modeling, as illustrated in the graphical abstract (Figure 2). The ANN model used a multilayer perceptron architecture comprising input, hidden, and output neurons, with two bias nodes included to facilitate adjustments (Figure 1D). The six variables identified by the RSF model were employed as input neurons. To optimize model performance, three hidden layers were added, and hyperparameters were fine-tuned using a five-fold cross-validation on the training dataset (Supplementary Table 1).

Figure 1 Modeling process. A: Top ten most important predictors based on random survival forests (RSF) analysis: Importance scores; B: Error rate and out-of-bag variable importance ranking from the RSF analysis; C: Predictors based on RSF analysis; D: Construction and specific structure of the artificial neural networks model. LMR: Lymphocyte-to-monocyte ratio; PLR: Platelet-to-lymphocyte ratio; ALBI: Albumin-bilirubin; AFP: Alpha-Fetoprotein; INR: International normalized ratio.

Figure 2 Graphical abstract patient enrollment and machine learning modeling process. TACE: Barcelona Clinic Liver Cancer Staging; RFA: Radiofrequency ablation; BCLC: Trans arterial chemoembolization.

Validation of the ANN model

As depicted in Figure 3A and B, in the training cohort, the AUC values of the ANN model at 1 year, 3 years, and 5 years were 0.92 (95%CI: 0.88-0.95), 0.87 (95%CI: 0.84-0.91), and 0.85 (95%CI: 0.82-0.88), respectively. Additionally, in the validation cohort, the AUC values of the ANN model were 0.81 (95%CI: 0.71-0.91), 0.79 (95%CI: 0.71-0.87), and 0.82 (95%CI: 0.74-0.85), respectively. Calibration curves demonstrated strong alignment between the predicted and actual survival probabilities (Figure 3C and D). The ANN model outperformed classic liver cancer models, such as the Okuda staging system (Okuda)[21], Japanese Integrated Staging (JIS)[22], Chinese University Prognostic Index (CUPI)[23], and Cancer of the Liver Italian Program (CLIP)[24] (DeLong test P < 0.05). The AUC values for each model are summarized in Table 2. Finally, the DCA curves demonstrated that the ANN model provided the highest net clinical benefit and exhibited stronger clinical utility compared to other classic liver cancer models (Figure 3E and F).

Figure 3 Evaluation of the predictive performance of the artificial neural networks model. A and B: The area under the curve of the artificial neural networks in prediction of 1-, 3-, and 5-year mortality in the training cohort (A) and validation cohort (B); C and D: Calibration curves for predicting 5-year overall survival in the training cohort (C) and validation cohort (D); E and F: The decision curve analysis for predicting 5-year mortality in the training cohort (E) and validation cohort (F). ROC: Receiver operating characteristic; OS: Overall survival; RSF: Random forest analysis. ANN: Artificial neural network; RSF: Random survival forest; JIS: Japan integrated staging score; Okuda: Okuda staging system; CUPI: Chinese university prognostic index; CLIP: Cancer of the liver Italian program.

Table 2 Comparison of the performance and discriminative ability between the artificial neural networks model and conventional models.

Corhort	Models	1-year AUROC (95%CI)	3-year AUROC (95%CI)	5-year AUROC (95%CI)	C-index
Training	ANN	0.92 (0.88-0.95)	0.87 (0.84-0.91)	0.85 (0.82-0.88)	0.81
	JIS	0.77 (0.70-0.83)	0.71 (0.65-0.78)	0.65 (0.61-0.70)	0.63
	Okuda	0.61 (0.53-0.69)	0.70 (0.65-0.75)	0.66 (0.62-0.70)	0.64
	CUPI	0.75 (0.66-0.83)	0.71 (0.63-0.87)	0.73 (0.69-0.77)	0.68
	CLIP	0.71 (0.63-0.78)	0.73 (0.68-0.78)	0.70 (0.66-0.74)	0.66
Validation	ANN	0.81 (0.71-0.91)	0.79 (0.71-0.87)	0.82 (0.74-0.85)	0.78
	JIS	0.60 (0.50-0.73)	0.66 (0.56-0.76)	0.61 (0.54-0.68)	0.61
	Okuda	0.67 (0.55-0.80)	0.69 (0.60-0.79)	0.69 (0.63-0.76)	0.63
	CUPI	0.78 (0.71-0.85)	0.75 (0.69-0.82)	0.76 (0.70-0.82)	0.68
	CLIP	0.70 (0.59-0.81)	0.71 (0.63-0.79)	0.72 (0.66-0.78)	0.66

ANN: Artificial neural network; RSF: Random survival forest; JIS: Japan integrated staging score; Okuda: Okuda staging system; CUPI: Chinese university prognostic index; CLIP: Cancer of the liver Italian program; AUROC: Area under the receiver operating characteristic.

Patient risk stratification

Based on the optimal cutoff value of the ANN model, patients were classified into low- and high-risk groups. The 5-year mortality rate for the high-risk group was 53.3% in the training cohort and 47.46% in the validation cohort (Figure 4A and B). Sub-group analysis showed that the ANN model maintained stability and provided satisfactory discriminatory ability for predicting 5-year prognosis among patients with early-stage HCC with different etiologies (with or without HBV infection), age (age < 60 years or age ≥ 60 years), and AFP levels (AFP < 400 ng/mL or AFP ≥ 400 ng/mL) (P < 0.001; Figure 4C-H). In Supplementary Figure 1, we also plotted the Kaplan-Meier curves of the other models for comparison. The results indicate that the ANN model consistently demonstrated superior stratification performance.

Figure 4 Kaplan-Meier curve analysis showing the efficacy of artificial neural networks levels as a predictor of 5-year mortality in patients with Barcelona Clinic Liver Cancer 0-A hepatocellular carcinoma after surgical resection across training and validation sets, different age groups, etiologies, and alpha-fetoprotein levels in the entire cohort. A: Kaplan-Meier survival curves for the training cohort; B: Validation cohort; C: Age < 60 years; D: Age ≥ 60 years; E: Patients with hepatitis B virus (HBV) infection; F: Patients without HBV infection; G: Alpha-fetoprotein (AFP) < 400 ng/mL; H: AFP ≥ 400 ng/mL. P values were obtained using the log-rank test. AFP: Alpha-fetoprotein; HBV: Hepatitis B virus.

Prognostic value of immune-inflammatory biomarkers

We compared immune-inflammatory markers between the survival and mortality groups among patients with early-stage HCC after surgical resection. The results showed that the PLR was elevated in the mortality group compared to that in the survival group, whereas the LMR was decreased (P < 0.05; Figure 5A and B). To validate the importance of immune markers in predicting the 5-year OS, we constructed a decision tree model incorporating six independent risk factors identified using the RSF model. The decision tree model stratified the variables into multiple levels based on their contribution to the outcomes (Figure 5C). As significant immune markers, the LMR and PLR were positioned at the first and fifth layers of the decision tree. Additionally, to enhance the interpretability of the ANN model, we conducted an evaluation using SHAP analysis. The SHAP Bar Plot and SHAP Summary Plot illustrated the overall impact of each feature on the model's predictions, with LMR and PLR showing significant positive and negative influences, respectively (Figure 5D and E). This ML validation further underscores the importance of immune markers in clinical management, suggesting their potential association with poor long-term prognosis in early-stage HCC patients following surgery.

Figure 5 Prognostic significance of immune-inflammatory biomarkers in patients with hepatocellular carcinoma. A: The levels of lymphocyte-to-monocyte ratio exhibited differences between patients in the survival and non-survival groups; B: The levels of platelet-to-lymphocyte ratio in the two groups; C: The decision tree model shows the 5-year overall survival probability of patients with hepatocellular carcinoma; D: The SHAP analysis displays the ranked importance of each feature in contributing to the overall predictions of the artificial neural networks (ANN) model; E: The SHAP analysis illustrates both the direction and magnitude of each feature's impact on the ANN model's predictions. LMR: Lymphocyte-to-monocyte ratio; ALBI: Albumin-bilirubin; PLR: Platelet-to-lymphocyte ratio; AFP: Alpha-Fetoprotein; INR: International normalized ratio.

DISCUSSION

For early-stage HCC, curative treatments such as liver transplantation or surgical resection should be chosen whenever possible, with a 5-year survival rate of around 60% for patients receiving these treatments[25]. However, even after surgical resection, the recurrence rate of HCC remains as high as 70%[26], highlighting the need for aggressive intervention in patients with early-stage liver cancer. Currently, imaging techniques, such as endoscopic ultrasound, computed tomography, and magnetic resonance imaging are primarily used to assess liver tumor prognosis. However, these methods are expensive and potentially harmful to the patients. To address this, we developed an ANN model based on demographic and common biological indicators for patients with early-stage liver cancer. This model integrates immune markers, tumor-related indicators, age, and liver function assessment scores to assist clinicians in identifying high-risk patients and improving long-term survival outcomes.

Traditional liver cancer models, including the Okuda, JIS, CUPI, and CLIP staging systems, have limitations. The Okuda staging system primarily considers whether the tumor volume is less than or greater than 50% of the liver volume[27], which may not be suitable for most patients with smaller tumors at an early stage. The JIS and CUPI systems mainly rely on the tumor, node, and metastasis criteria[28], which focus on tumor characteristics and has low efficiency in predicting early-stage liver cancer outcomes owing to the subtlety of tumor characteristics at this stage. Additionally, CLIP staging is based on European patient cohorts, but the etiology of liver cancer differs between developed and developing countries, such as China, where HBV infection is more prevalent. In our cohort, 80% of liver cancer patients tested positive for the HBV surface anti-gen, whereas in developed European regions, liver cancer is usually caused by alcoholic or fatty hepatitis[29]. Therefore, developing a more specialized model to forecast the prognosis of Asian patients with BCLC 0-A stage liver cancer is necessary. In this study, our new ML model integrated biomarkers with hidden layers, effectively weighting a wide range of variables using a black-box algorithm. These models outperform traditional models in predictive performance and better manage the nonlinear relationships among indicators[30]. By utilizing the ANN model, we can more accurately identify high-risk patients, thereby optimizing patient stratification. This more precise stratification will help in developing personalized treatment plans for patients at different risk levels, enabling earlier and more aggressive interventions and ultimately improving long-term survival rates.

Previous studies have demonstrated that inflammation can contribute to the formation of a complex tumor microenvironment in the presence of immune cells. Monocytes can induce immune tolerance, promote angiogenesis, and facilitate tumor cell dis-semination[31,32]. Neutrophils can locally promote primary tumor growth and metastasis[33]. Platelets, non-hematopoietic components responsible for the inflammatory response, release various proangiogenic proteins to aid tumor growth[34] and produce cytokines and chemokines to form the tumor microenvironment[35]. Although the activation and exhaustion of these inflammation-related immune cell markers are often not directly reflected by an increase or decrease in their quantities, their ratios can provide a preliminary indication. Previous studies have revealed that patients with an elevated baseline NLR often have a high risk of recurrence and poor efficacy of targeted therapies following liver transplantations[36]. Similarly, a meta-analysis showed that a high preoperative PLR increases the risk of postoperative recurrence by approximately 3.33 times, and that high PLR levels were closely associated with extrahepatic metastasis of the tumor[37]. Additionally, in 2014, Hu et al[38] developed the systemic immune-inflammation index (SII), which combines lymphocyte, neutrophil, and platelet counts. The SII was found to be an independent prognostic factor for postoperative outcomes in patients with HCC[38]. In our study, we also found that compared to the survival group, patients in the mortality group had elevated PLR levels, while LMR levels decreased. According to the decision tree model, patients with LMR < 2.2 had a risk ratio of approximately 2.3, making it the most important immune-inflammatory diagnostic marker.

ANNs and RSFs are among the most commonly used ML methods, and both are capable of constructing nonlinear statistical models to achieve high-precision predictions. In particular, ANNs can minimize errors through self-learning adjustments of the weights between the inputs and outputs. In this study, the AUC values of the ANN model for predicting 1- to 5-year survival in patients with HCC were considerably greater than those of the traditional scoring systems, such as the Okuda, JIS, CUPI, and CLIP methods. Based on its cutoff values, the ANN model demonstrated excellent discriminative ability for patients with varying levels of liver function and disease etiologies. According to the cutoff value, the hazard ratio for high-risk vs low-risk patients in the training cohort was 7.98 (95%CI: 5.85-10.93), whereas in the validation cohort, it was 6.74 (95%CI: 4.23-10.69). The low cutoff value of the ANN model achieved a negative predictive value of 91.37% in the training cohort and 90.90% in the validation cohort, excluding patients with a good 5-year prognosis. We also noted that the DCA curves showed minimal differences in net benefit among the ANN, CLIP, and CUPI models at lower thresholds. However, when the threshold probability was below 0.35, the primary aim of the predictive model was to minimize missed diagnoses (false negatives), thereby offering early intervention opportunities for more potentially high-risk patients. In this range, the ANN model exhibited stable performance, providing enhanced decision support for clinicians and aiding in the identification of patients who might be overlooked by traditional scoring systems. Furthermore, decision-tree nodes, as part of the ML algorithm, indicated the method of variable splitting, with the hierarchy representing the contributions of different predictive variables to the outcome. Our results showed that the LMR and PLR exhibited significant differences between the deceased and surviving groups, underscoring their critical impact on long-term outcomes and highlighting the importance of monitoring immune markers in clinical practice. These findings underscore the potential of ANN models in personalized medicine, particularly for prognostic assessment and treatment decision-making for patients.

Finally, we acknowledge the limitations of this study: First, this study was based on retrospective data from a single-center sample, and thus selection bias may have occurred. Future research should include multi-center and multi-regional data, to ensure the generalizability of the model. Second, in China, 80% of liver cancer cases are caused by HBV infection. Therefore, the model may exhibit bias for patients with different etiologies. Future research should validate the predictive capability of the model across different etiological backgrounds. Third, this study was based on a static, retrospective dataset, and only 5-year survival data were tracked. Continuous follow-up of patient survival data is necessary to further refine and expand the application of the ANN model. Fourth, retrospective study data inherently carry issues, such as selection and confounding bias. In future studies, data should be collected from more comprehensive sources using standardized, well-developed data collection methods. We aim to promote our model across major medical institutions and expand the research cohorts to enhance confidence in and familiarity with the model. At the same time, we will conduct a prospective multi-center study to identify the most effective follow-up treatment strategies for high-risk patients after surgery, with the goal of improving survival rates and providing stronger decision-making support for clinicians.

CONCLUSION

This study developed a predictive model using ML algorithms that integrates immune-inflammatory biomarkers to predict the long-term prognosis of patients with early-stage HCC after surgical resection, thereby aiding the optimization of screening and treatment strategies. However, future large-scale, multi-center prospective studies are needed to validate the performance of the model.

ACKNOWLEDGEMENTS

We sincerely thank all the patients and their families for their support of this study.