A novel nomogram based on machine learning predicting overall survival for hepatocellular carcinoma patients with dynamic α‑fetoprotein level changes after local resection

Introduction

Primary liver cancer has emerged as a substantial economic burden in global public health, characterized by its high incidence and mortality rates (1). In these cases, hepatocellular carcinoma (HCC) is the predominant form, accounting for 75–80% (2). Notably, China bears a disproportionately high incidence of HCC, contributing to >50% of cases worldwide (3). Currently, there are various treatment options for HCC, including surgical resection, ablation, transarterial chemoembolization (TACE), liver transplantation, systemic therapy and combination therapy (4–6). Among these options, liver resection is the preferred choice for patients with well-functioning livers and localized tumors, offering the potential for cure (7). However, in most Asian centers, overall survival (OS) rates for patients remain unsatisfactory (8). Approximately half of all HCC patients experience recurrence or metastasis within 5 years after curative resection (9,10). Therefore, it is imperative to identify risk factors that impact OS in HCC patients after liver resection.

α-fetoprotein (AFP), as a pivotal biomarker in the detection of HCC, plays an essential role in accurate diagnosis, early identification, evaluation of treatment effectiveness, recurrence monitoring and prognosis prediction for HCC (11). Nonetheless, it is noteworthy that ~30% of HCC patients do not exhibit elevated serum AFP levels (12). The delayed diagnosis of AFP-negative HCC (AFP-NHCC) often results in treatment postponements, affecting patient prognosis (13). Even during postoperative follow-up, patients with AFP-NHCC continue to undergo regularly AFP level monitoring, given the close correlation between AFP and tumor prognosis (14,15). The findings of the present study indicated that certain patients with AFP-NHCC experience dynamic changes in AFP levels after resection. While current studies have focused on identifying prognostic factors in patients with AFP-NHCC after hepatectomy. A consensus on the prognostic factors for patients experiencing postoperative AFP level elevation remains elusive and research on the risk factors related to OS for this patients cohort is lacking.

The Random Survival Forest (RSF) is a powerful machine-learning algorithm composed of multiple decision trees, demonstrating relatively high accuracy, robustness and strong resistance to over-fitting. Therefore, by combining the RSF with traditional multivariate Cox regression, more reliable prognostic factors related to OS can be identified and a nomogram can be constructed using the aforementioned variables. A nomogram is a common visual representation of clinical prediction models. Constructed from a comprehensive combination of multiple clinical indicators, the nomogram enables physicians to deliver more direct and accurate prognoses for specific patients. This assists in adjusting treatments accordingly, aiming for improved clinical outcomes. Consequently, the present study aimed to predict the OS of HCC patients who initially tested negative for AFP at baseline but exhibited a subsequent change to AFP positivity during follow-up after curative resection. This prediction was based on clinical data and provides improved guidance for patient management.

Materials and methods

Patients

The present study retrospectively analyzed 870 HCC patients who underwent resection at Beijing You'an Hospital, Capital Medical University, China between January 2013 and January 2021. The age of all patients ranged from 22 to 78 years, with a mean age of 56.61±9.08 years, and the male-to-female ratio was 3.58:1. The cohort was randomly divided into a training set and a validation set at a ratio of 7:3. The training set was used for variable selection and model construction, while the validation set was used to confirm the performance of the developed model. Baseline characteristic in the training set and verify set are given in Table I.

Table I. Baseline characteristic in the training set and validation set.

Table I.

Baseline characteristic in the training set and validation set.

Variable	Group	Training set (n=600)	Validation set (n=270)	P-value
Age, n (%)	≤60 years	373 (62.17)	168 (62.22)	0.988
	>60 years	227 (37.83)	102 (37.78)
Sex, n (%)	Male	469 (78.17)	211 (78.15)	0.995
	Female	131(21.83)	59 (21.85)
Hypertension, n (%)	Yes	162 (27.00)	72 (26.67)	0.918
	No	438 (73.00)	198 (73.33)
Diabetes, n (%)	Yes	140 (23.33)	52 (19.26)	0.180
	No	460 (76.67)	218 (80.74)
Smoking, n (%)	Yes	251 (41.82)	119 (44.07)	0.536
	No	349 (58.17)	151 (55.93)
Drinking, n (%)	Yes	189 (31.50)	94 (34.81)	0.334
	No	411 (68.50)	176 (65.19)
Cirrhosis, n (%)	Yes	514 (85.67)	238 (88.15)	0.323
	No	86 (14.33)	32 (11.85)
Child-Pugh, n (%)	A	462 (77.00)	196 (72.59)	0.161
	B	138 (23.00)	74 (27.41)
BCLC, n (%)	0	233 (38.83)	95 (35.19)	0.304
	A	367 (61.17)	175 (64.81)
Tumor number, n (%)	Single	498 (83.00)	210 (77.78)	0.067
	Multiple	102 (17%)	60 (22.22)
Tumor size, n (%)	≤3 cm	443 (73.83)	198 (73.33)	0.877
	>3 cm	157 (26.17)	72 (26.67)
RBC (mean ± SD), 1012/l	-	4.18±0.63	4.14±0.60	0.376
HB (mean ± SD), g/l	-	131.51±19.73	129.97±18.77	0.280
WBC (mean ± SD), 109/l	-	5.12±2.16	4.88±1.98	0.113
AST (mean ± SD), IU/l	-	31.38±15.00	33.07±16.54	0.136
ALT (mean ± SD), IU/l	-	30.57±18.82	33.29±22.17	0.063
TBIL (mean ± SD), µmol/l	-	19.45±10.06	19.42±10.28	0.975
DBIL (mean ± SD), µmol/l	-	6.66±4.66	6.69±4.97	0.910
Total Protein (mean ± SD), g/l	-	65.37±6.80	65.21±8.23	0.774
GGT (mean ± SD), U/l	-	63.97±57.50	68.64±54.33	0.260
ALP (mean ± SD), IU/l	-	86.98±35.86	87.95±34.59	0.711
Glob (mean ± SD), g/l	-	28.08±5.37	28.25±5.05	0.657
PT (mean ± SD), s	-	12.60±1.57	12.53±1.44	0.526
PTA (mean ± SD), %	-	85.94±15.12	86.68±14.60	0.502
INR (mean ± SD)	-	1.12±0.14	1.11±0.13	0.359
Fibrinogen (mean ± SD), mg/dl	-	2.78±0.92	2.76±0.91	0.827

[i] Continuous variables were presented as mean ± standard deviation. Categorical variables were described as frequency and percentage. BCLC stages, Barcelona Clinic Liver Cancer stages; RBC, red blood cell; HB, Hemoglobin; WBC, white blood cell; AST, aspartate aminotransferase; ALT, alanine aminotransferase; TBIL, total bilirubin; TBIL, direct bilirubin; GGT, gamma-glutamyl transferase; ALP, alkaline phosphatase; Glob, globulin; PT, prothrombin time; PTA, prothrombin activity; INR, international normalized ratio.

The present study was conducted according to the Declaration of Helsinki and approved by the Ethics Committee of Beijing You'an Hospital, Capital Medical University, China (approval no. LL-2021-152-K) and followed the 1964 Helsinki Declaration and its later amendments or comparable ethical standards.

The inclusion criteria for the present study were as follows: i) Pathological diagnosis of HCC, ii) Surgical resection of single or multiple isolated HCC lesions in one liver lobe, iii) Preoperative serum AFP level <20 ng/ml. During follow-up (>2 months after resection), patients exhibited elevation of AFP level and iv) Early-stage HCC patients. Exclusion criteria included: i) Previous receipt of other treatments such as ablation and TACE ii) concomitant other malignancies and iii) missing information.

Data collection

Prior to patients undergoing HCC resection, baseline data and tumor characteristics of all patients were obtained from the medical records system, including age, sex, medical history, family history, presence of cirrhosis, Barcelona Clinic Liver Cancer (BCLC) staging, Child-Pugh classification (16), tumor number and size, AFP levels, as well as routine blood test parameters, liver function indices and coagulation indicators. The outcome of resection was primarily obtained through telephone follow-up. The postoperative AFP levels were mainly retrieved from the electronic medical records when patients are readmitted to the hospital.

Follow-up and endpoint

All patients underwent curative liver resection after completing preoperative evaluations. Postoperatively, patients were regularly followed up with various assessments, including physical examinations and monitoring of serum AFP levels. The clinical endpoint was OS, with follow-up data collected until January 2023. OS was defined as the time between the first liver cancer resection and either death or the last follow-up date.

Statistical analysis

Data analysis used IBM SPSS Statistics 27 (IBM Corp.) and R version 4.1.2 (17). Normally distributed data were presented as mean ± standard deviation, while skewed data was represented using quartiles. The Mann-Whitney U test or Student's t-test was employed to compare numerical variables between the derivation set and internal validation set, while Fisher's exact test or χ2 test was used for assessing categorical variables. Significant predictive factors were identified using RSF and multivariate Cox regression. These variables were subsequently incorporated into a nomogram model. Based on a total score of 130, patients were assigned to the low, medium and high-risk groups using cutoff values of 43 and 86. Kaplan-Meier (KM) analysis and log-rank tests were used to compare differences in OS among the three different risk groups. The areas under the curves (AUCs) of the time-dependent receiver operating characteristic (ROCs) curves were employed to evaluate model accuracy. Additionally, the concordance index (C-index) was used to assess model discrimination, while calibration curves were employed to evaluate consistency between training and validation sets. Clinical utility assessment was conducted through decision curve analysis (DCA), quantifying net benefits at different probability thresholds.

Results

Baseline characteristics and survival outcomes of patients

The final analysis involved 870 eligible patients with AFP-NHCC, divided into the model development group (n=600) and the validation group (n=270). Baseline data comprised 26 items. In the entire cohort, 62.18% of patients were <60 years old and 78.16% were male. Additionally, 86.44% of patients had a background of cirrhosis, with the majority falling into Child-Pugh A (75.63%) and 62.30% of patients were in BCLC stage A. No statistical differences were observed between the model development and validation groups (all P>0.05), indicating consistency between the two groups. Detailed demographic and clinical characteristics between the model development group and the validation group are presented in Table I.

The 3-, 5- and 8-year overall survival rates for all patients were 90.3, 76.7 and 60.5%, respectively. The KM curve of overall survival for the patients is shown in Fig. S1.

Identification of predictors and development of the nomogram

The variables in the model-building group were analyzed, tested and adjusted. Finally, RSF identified the top five prognostic factors, including sex, tumor size, globulin, gamma-glutamyl transferase (GGT) and fibrinogen (Fig. 1). The reliability of the aforementioned variables was also confirmed by the multifactorial COX regression (Table II). The present study incorporated the aforementioned parameters to combine into a new model and visualized the model with nomogram (Fig. 2). By calculating the total scores, surgeons can easily obtain the probability of OS as predicted by the nomogram.

Figure 1. The RFS model for variables. (A) Error rate when number of trees=530. (B) Variable importance for each variable. RSF, random survival forest; Glob, globulin; GGT, gamma-glutamyl transferase; Fib, fibrinogen; T.S., tumor size.

Figure 2. Nomogram, including sex, Glob, T.S., Fib and GGT for 3-, 5- and 8-years OS in patients with AFP-NHCC with dynamic AFP level changes. Glob, globulin; GGT, gamma-glutamyl transferase; AFPN-HCC, α-fetoprotein-negative hepatocellular carcinoma; Fib, fibrinogen; T.S., tumor size; OS, overall survival.

Table II. Multivariate Cox regression analysis of risk factors for OS in the training cohort.

Table II.

Multivariate Cox regression analysis of risk factors for OS in the training cohort.

Variable	P-value	HR	95%CI
Sex	0.004	0.653	0.491–0.869
Tumor size	0.018	1.919	1.659–3.282
Glob	0.015	1.431	1.073–1.729
GGT	0.032	1.671	1.098–1.983
Fibrinogen	0.035	1.114	1.095–1.306

[i] Glob, Globulin; GGT, gamma-glutamyl transferase; HR, hazard ratio; CI, confidence interval.

Evaluation of the nomogram

The predictive performance of the nomogram was assessed using the C-index, revealing a C-index of 0.72 (95% CI: 0.685–0.755) for the model-development group. ROC curve analysis demonstrated area AUCs of 0.731, 0.748 and 0.775 at 3, 5 and 8 years, respectively (Fig. 3A). Calibration curves illustrated good consistency between predicted and actual values for the 3-, 5- and 8-year OS rates (Fig. 4A-C). DCA further indicated the nomogram as a valuable predictive tool (Fig. 5A-C). Moreover, stratifying patients into high-risk, intermediate-risk and low-risk groups showed significant statistical differences in survival rates in the training set (P<0.001; Fig. 6A).

Figure 3. Time-dependent ROC curves for the nomograms in the test and proof cohorts. (A) The AUCs for OS at 3, 5 and 8 years in the test cohort. (B) The AUCs for OS at 3, 5 and 8 years in the proof cohort. ROC, receiver operating characteristics; AUC, area under the curve; OS, overall survival.

Figure 4. Calibration curve of the nomogram in the training and validation cohort. (A) 3-year OS in the test cohort. (B) 5-year OS in the test cohort. (C) 8-year OS in the test cohort. (D) 3-year OS in the proof cohort. (E) 5-year OS in the proof cohort. (F) 8-year OS in the proof cohort. OS, overall survival.

Figure 5. DCA for nomogram depicts the clinical net benefit. (A) 3-year DCA in the test cohort. (B) 5-year DCA in the test cohort. (C) 8-year DCA in the test cohort. (D) 3-year DCA in the proof cohort. (E) 5-year DCA in the proof cohort. (F) 8-year DCA in the proof cohort. DCA, decision curve analysis.

Figure 6. Kaplan-Meier survival curves for patients with high-, intermediate- and low-risk by the nomogram score. (A) Test cohort. (B) Proof cohort.

Validation of the nomogram

To ascertain the reliability of the nomogram, internal validation process was undertaken. The C-index for the validation cohort was 0.664 (95% CI: 0.611–0.717), indicating that the predicted results of the nomogram were consistent with the actual observed outcomes, confirming the good discriminative ability of the model. The corresponding AUCs for 3-, 5- and 8-year ROC curves were 0.664, 0.708 and 0.753, respectively, attesting to its strong classification capability (Fig. 3B). The calibration curves also exhibited good alignment between predicted and observed values (Fig. 4D-F). DCA curves validated good clinical utility as well as a fine balance between benefits and risks (Fig. 5D-F). Lastly, the OS rates stratified by risk groups were consistent with those in the model development group, also showing notable disparities (P<0.01; Fig. 6B).

Discussion

Approximately 30% of HCC cases are AFP-NHCC, representing a clinically distinct subgroup that is frequently underdiagnosed owing to nonspecific clinical manifestations. Consequently, the delayed diagnosis and treatment poses threat to their survival (12,18). Among these patients, some exhibit dynamic changes from AFP negativity to AFP positivity indicating a unique population within patients with AFP-NHCC. AFP levels have been associated with the pathological grading, progression and prognosis of the patients, suggesting that this subgroup of patients differ from AFP-negative patients (14,15). After surgical resection for HCC, patients may sustain varying degrees of damage to both the tumor and its surrounding tissues, leading to inflammation. Additionally, the liver initiates a regenerative process to repair the damaged areas following the resection (19,20). These responses can both stimulate hepatocyte proliferation and the secretion of AFP, ultimately resulting in a temporary elevation of AFP levels post-surgery. However, this is not indicative of a poor prognosis. Furthermore, guidelines indicate that for AFP-positive liver cancer patients, AFP levels typically return to normal within two months after hepatic resection (21). To ensure that fluctuations in AFP levels in AFP-negative HCC patients after hepatic resection were not transient increases, the present study established a two-month observation period as a basis for assessing AFP level fluctuations. AFP levels, to a certain extent, reflect the size of the tumor and their dynamic changes are related to the disease status, serving as a sensitive indicator for assessing treatment efficacy and prognosis (21–24). Additionally, early-stage liver cancer or liver cancer confined to a single lobe, with small lesions and no metastasis, generally responds well to treatment. By contrast, advanced liver cancer, characterized by larger lesions often accompanied by metastasis, presents markedly greater treatment challenges. Despite various treatment options such as immunotherapy and targeted therapy, overall treatment outcomes remain limited and patient survival rates are unsatisfactory (25–28). Survival predictions for early-stage liver cancer can help doctors devise more reasonable treatment plans and improve treatment outcomes. However, for patients with advanced liver cancer, even with survival predictions, the rapid progression of the disease and poor treatment efficacy may lead to significant discrepancies between predicted and actual survival outcomes. Furthermore, clinical practice focusses more on early detection of disease changes, enhancing treatment efficacy and improving patient prognosis. Therefore, research on survival predictions for early-stage liver cancer confined to a single lobe aligns more closely with the needs of clinical practice. Therefore, the present study aimed to focus on this special patient cohort and develop a nomogram to predict their OS. Moreover, for these patients, treatments such as liver transplantation and extended resection should be considered and follow-up frequency should be increased (29,30).

The nomogram developed in the present study incorporated sex, tumor size, globulin, GGT and fibrinogen. It assigned scores to each level of these key factors based on their contribution to the outcome variable within the model. By aggregating these scores, the total score was then translated into the predicted OS probability using a function transformation relationship. The model demonstrated robust predictive capability through multidimensional validation. As evidenced by the calibration curve and ROC analysis, the model displayed robust diagnostic performance. DCA analysis also indicated satisfactory predictive ability, while the variable availability made this model user-friendly for practical clinical applications. KM analysis further confirmed the ability of this model in clinical practice. Stratifying patients based on total scores into low, intermediate and high-risk groups revealed markedly distinct OS rates.

Previous studies (31–35) have primarily focused on the prognosis of AFP-NHCC patients after hepatectomy, neglecting the dynamic changes in AFP levels. To address this gap, the present study developed a more accurate predictive model for this subgroup to improve their prognosis. The cohort was not only massive in scale but also had a long follow-up period, lending greater credibility to the conclusions. In addition, published reports (36–40) have shown the trustworthiness of the aforementioned five indicators in predicting OS for patients with HCC. Generally, male patients have poorer prognoses. Beyond liver cancer, they are more prone to non-reproductive system tumors and worse prognoses, possibly due to factors related to Y chromosome genes and testosterone (28,41,42). Studies suggest that testosterone promotes CD8+ T cell exhaustion, leading to faster tumor cell growth (43,44). Additionally, another study indicates a correlation between higher levels of Inc-FTX, a regulator transcribed from the X chromosome inactivation center XIST, and longer prognosis in HCC patients. Inc-FTX acts as a tumor suppressor, with higher expression levels observed in the female liver (45). Tumor size is a critical determinant of the 2-year postoperative recurrence rate in isolated HCC. Specially, patients with a tumor diameter >5 cm and AFP ≥20 ng/ml have a 4.5 times higher mortality rate than those with a tumor diameter <5 cm (46,47). Retrospective analysis of patients surviving postoperative HCC for over 10 years found that isolated and small tumors are critical factors for long-term postoperative survival (48). The globulin levels in HCC patients are markedly elevated compared with healthy individuals. As HCC continues to progress, it can activate the body's immune mechanism, leading to the production of a large number of inflammatory factors, further burdening the liver and affecting its normal physiological functions. This leads to a further increase in globulin levels, which in turn affects the patient's liver function, thereby exacerbating HCC. GGT, a key enzyme associated with liver metabolism, has recently been implicated in oxidative stress, extracellular inflammation and tumor progression (49). Inflammation stimuli within or surrounding HCC may induce abundant GGT production in hepatocytes and the cancer cells themselves also synthesize GGT, further increasing serum GGT levels. Moreira et al (50) demonstrated that serum GGT levels increase with the progression of liver cancer and promote tumor advancement in the male Wistar rat HCC animal model. Moreover, GGT levels are closely linked to the prognosis of patients with AFP-NHCC (51–53). AFP-low-level HCC patients, those with high GGT levels, are more likely to experience lower survival rates (54), in accordance with the findings of the present study. Fibrinogen, a common coagulation-related protein, apart from participating in blood clotting, is closely associated with tumors (55). Studies (56–60) reveal a significant increase in preoperative plasma fibrinogen in various malignant tumors, closely correlating with tumor progression, metastasis and prognosis. Elevated fibrinogen usually implies a hypercoagulable and inflammatory state, which inevitably affects the patient's postoperative recovery and prognosis, prolongs the postoperative hospital stay and affects the OS of the patient.

Limitations existed in the present study. First, retrospective studies inevitably introduce selection bias, mitigated to some extent by the ample sample size of the present study. Second, although the model exhibited good internal performance during validation, external validation in additional cohorts is necessary to enhance the credibility and persuasiveness of the model, as well as to validate its generalization ability. Third, the present study did not include recurrence patterns or surgery-related indicators, such as the extent of operation, surgical margin, need of blood transfusion and tumor pathology. More comprehensive indicators should be encompassed in any future study. Nevertheless, the model still provided more timely treatment guidance for patients with AFP-NHCC who exhibit dynamic AFP changes due to the identification of high-risk patients in this subgroup.

Supplementary Material

Supporting Data

Acknowledgements

Not applicable.

Funding

The present study was funded by the National Science and Technology Major Project (grant nos. 2023ZD0502405 and 2023ZD0502402) and 2023 Young and middle-aged Talents Incubation Project (Youth Innovation) of Beijing You'an Hospital, Capital Medical University (grant no. BJYAYY-YN2023-13).

Availability of data and materials

The data generated in the present study may be requested from the corresponding author.

Authors' contributions

Conceptualization and design were performed by RJ and GL. Data collection was performed by QW, LS, GZ and ZC. Data analysis and interpretation were conducted by QW, LS and GZ. The manuscript was drafted by QW, ZC and LS. Critical review and editing were carried out by RJ and GL. GL obtained funding. Supervision was provided by RJ and GL. GL and RJ confirm the authenticity of all the raw data. All authors read and approved the final version of the manuscript.

Ethics approval and consent to participate

The present study was conducted according to the Declaration of Helsinki and approved by the Ethics Committee of Beijing You'an Hospital, Capital Medical University, China (approval no. LL-2021-152-K) and followed the 1964 Helsinki Declaration and its later amendments or comparable ethical standards. The Institutional Review Board of Beijing You'an Hospital, Capital Medical University, China waived informed consent because of the retrospective nature of our study.

Patient consent for publication

Not applicable.

Competing interests

The authors declare that they have no competing interests

Glossary

Abbreviations

Abbreviations:

References