Machine learning model-based approach using cellular proliferation marker expression for preoperative clinical decision-making in patients with hepatocellular carcinoma

Abstract

The investigation by Zhu et al on the assessment of cellular proliferation markers to assist clinical decision-making in patients with hepatocellular carcinoma (HCC) using a machine learning model-based approach is a scientific approach. This study looked into the possibilities of using a Ki-67 (a marker for cell proliferation) expression-based machine learning model to help doctors make decisions about treatment options for patients with HCC before surgery. The study used reconstructed tomography images of 164 patients with confirmed HCC from the intratumoral and peritumoral regions. The features were chosen using various statistical methods, including least absolute shrinkage and selection operator regression. Also, a nomogram was made using Radscore and clinical risk factors. It was tested for its ability to predict receiver operating characteristic curves and calibration curves, and its clinical benefits were found using decision curve analysis. The calibration curve demonstrated excellent consistency between predicted and actual probability, and the decision curve confirmed its clinical benefit. The proposed model is helpful for treating patients with HCC because the predicted and actual probabilities are very close to each other, as shown by the decision curve analysis. Further prospective studies are required, incorporating a multicenter and large sample size design, additional relevant exclusion criteria, information on tumors (size, number, and grade), and cancer stage to strengthen the clinical benefit in patients with HCC.

Key Words: Hepatocellular carcinoma; Machine learning model; Cellular proliferation marker; Preoperative therapy decision; Cancer

Core Tip: The retrospective study by Zhu et al employed a machine learning model to evaluate cellular proliferation markers in patients with hepatocellular carcinomas, demonstrating its predictive ability and clinical benefits in presurgery treatment decisions. Retrospective cancer prognostic biomarker studies face limitations such as selection bias, data quality, factors affecting biomarker-patient outcomes, and poor generalizability to different populations. The study was based on a small population with no geographical information in the report. The study lacks information on tumor histology (size, number of tumors, grade, and primary/secondary nature), which is highly associated with the marker signature in the samples.

TO THE EDITOR

I am delighted to read the high-quality article by Zhu et al[1]. The retrospective study aimed to develop a machine learning model for assessing intratumoral and peritumoral region markers in hepatocellular carcinoma (HCC) before surgery and aiding in clinical decision-making. Pathologically diagnosed patients with HCC with high (n = 104) and low cellular proliferation marker (Ki-67) (n = 64) expression were considered for the study. Radiomic features of the intratumoral, peritumoral, and fused volume of interests were selected by least absolute shrinkage and selection operator regression screening. Logistic regression, area under curve, and waterfall plot analysis demonstrated better performance of the fusion model in terms of higher accuracy in Ki-67 expression assessment. Univariate and multivariate logistic regression analysis found a positive association of Radscore of the fusion model (alanine aminotransferase, HBV infection, and cirrhosis as clinical features) with the increased Ki-67 expression. Further, the Radscore from the fusion model and clinical parameters from the clinical model were merged to create a clinical-radiomic nomogram. The calibration curve had good consistency between projected and real Ki-67 expression probability, and the decision curve proved its clinical utility in patients with HCC.

This report assessed the utility of a clinical-radiomic nomogram based on Ki-67 expression and its association with the clinical features in the intratumoral and peritumoral regions of patients with HCC. They discovered that the Ki-67 expression-based clinical-radiomic nomogram model helped doctors decide what treatment to use before surgery. Although their data is intriguing, I have some issues with the paper. The consideration of information about tumors is lacking in the study. Tumor size, number, grade, and cancer stage at diagnosis affect HCC prognosis. For example, larger tumors and advanced stages generally indicate a worse prognosis[2].

The potential for significant bias resulting from the inherent limitations of analyzing data collected in the past, such as inconsistent data collection practices, missing information, and selection bias, is a major drawback of a retrospective prognostic biomarker study in cancer. These limitations can have a significant impact on the validity and reliability of the conclusions of the study regarding the prognostic value of the biomarker[3]. Selection bias, data quality, factors affecting the association between the biomarker and patient outcome (therapy response and survival), and poor generalizability to other populations are some of the limitations of retrospective prognostic biomarker studies in cancer. Thus, a prospective study design with careful attention to the mentioned concerns may enhance its reliability, consistency, and generalizability to other populations of patients with HCC.

HCC ranks fourth in global cancer deaths and is the fastest-growing cancer-specific mortality in the United States. HCC often develops in patients with chronic liver disease with hepatoviral infection or nonalcoholic fatty liver disease[4]. Patients with HCC have an 18% 5-year survival rate[5]. Because most patients (60%) have advanced disease, curative purpose treatment is not possible. Thus, precise prognostic indicators may improve patient selection and identify those who benefit from aggressive HCC treatment.

Ki-67 is a protein located in the nucleolar cortex and is expressed in most proliferating malignant cells but rarely in normal cells. Ki-67 enters chromosomes during cell division and scales throughout grade 1 (G1) until mitosis before rapidly decreasing. It is one of the most often used clinical indicators for cell proliferation in several cancers[5]. High and low Ki-67 expression levels are crucial for prognosis and therapy considerations[6]. Previous research has linked Ki-67 expression to worse tumor biology and prognosis in patients with cancer. A recent meta-analysis found that patients with HCC with high Ki-67 protein expression had larger tumors, more lymph node metastases, cirrhosis, vascular invasion, and distant metastasis[5]. Paraffin-embedded slice immunohistochemistry measures the Ki-67 marker index for prognosis. Conventional imaging is unable to distinguish subtle differences between HCC with varying levels of Ki-67 expression[6].

Patients with low Ki-67 expression may benefit more from surgery as they are less likely to have early recurrence. High Ki-67 Levels can indicate higher recurrence risk, influencing transplant eligibility and post-transplant surveillance. High Ki-67 Levels may indicate a greater likelihood of response to targeted therapies (e.g., sorafenib, lenvatinib) and systemic chemotherapy. Sorafenib is a targeted kinase inhibitor that causes cancer cell autophagy, suppressing the growth of HCC; it is recommended either as adjuvant or neoadjuvant treatment. Among other applications, the described model could determine neoadjuvant sorafenib administration, particularly in those patients with HCC waiting on liver transplantation who meet the Milan criteria. This treatment while waiting for transplantation may avoid upstaging, thus preventing cancellation of the prospective transplantation. Thus, the potential use of the present Ki-67 based machine learning model should also be extended to clinical decisions in targeted therapy and chemotherapy in patients with HCC.

High Ki-67 expression in patients with HCC is linked to poor overall and disease-free survival, with a 2-3 times higher risk of recurrence post-surgery or liver transplantation[5,7]. HCC progression is influenced by a complex interplay of genetic/epigenetic/metabolic alterations, inflammatory responses, angiogenesis, hypoxia, signaling pathways, and therapy resistance. A longitudinal study on machine learning-based Ki-67 expression detection and patient outcomes can establish standardized cutoff values for stratifying patients with low-risk and high-risk HCC, improving risk assessment for post-surgical recurrence and long-term survival.

Further, the use of multi-biomarker models (including Ki-67) could potentially improve the predictive accuracy of treatment response, potentially enhancing patient selection for immunotherapy, targeted therapy, and combination therapies. Future research could explore machine learning-based Ki-67 expression and genetic mutations in HCC, enabling personalized therapies and patient-derived sample analysis for treatment changes, guiding adaptive treatment plans. Overall suggestions made in this article will contribute to a more profound understanding of machine-based learning of Ki-67 expression profile in HCC progression, treatment response, and personalized care. By improving risk stratification, optimizing therapy selection, and refining post-treatment monitoring, these suggestions can enhance clinical outcomes and survival rates in patients with HCC.

Imaging examinations of cancer tissue samples help oncologists to assess therapeutic efficacy and disease prognosis. The radiologist evaluates the imaging data sets and classifies the response using response evaluation criteria. More recently, radiologists have developed quantitative and objective estimates of CT to summarize and facilitate patient clinical treatment[8,9]. Hepatic lesions are extremely frequent in oncology[10]. Liver metastasis is prevalent, and HCC, the most common primary tumor, is the second greatest cause of mortality in oncological patients and the top in patients with cirrhosis.

Many researchers have investigated radiomics for liver lesions in recent years. Today, researchers have consolidated many applications, ranging from early diagnosis to post-treatment evaluation and prognosis predictions[11]. Radiomics have been used to predict the risk of recurrence in primary hepatic cancers, with studies showing that clinical factors (such as ferritin, macrovascular invasion, tumor size, and sex) and radiomic features can be considered as potential imaging biomarkers for patient survival[12,13].

Most hepatic lesions are metastases, which are 18-40 times more common than initial tumors. Hepatic metastases have been the subject of many radiomic investigations due to necessity. Radiomic characteristics can also predict treatment efficacy in patients with hepatic metastases[14,15]. Ravanelli et al[16] used contrast-enhanced CT to extract radiomic features, showing its importance in predicting overall survival and disease-free survival in patients with liver metastases from unresectable colorectal cancer. Nakanishi et al[17] developed a model to predict the response of liver metastases to first-line oxaliplatin-based chemotherapy in patients with colorectal cancer using radiomic features extracted from pretreatment CT scans. Thus, caution should be exercised about the information on the type of lesion (metastases or initial tumor) during the development of the machine learning-based model for patients with HCC.

CONCLUSION

This single-center retrospective study had limitations due to limited case numbers and reliance on CT images. It also lacked a universally accepted threshold for high Ki-67 expression in liver cancer. The peritumoral zone was established as 3 mm, but further research is needed. The study used machine learning methods, but deep learning could enhance the predictive ability and clinical value. Overall, a prospective study design is required by including a larger number of samples from multiple centers, information on tumor histopathological information, and broadening inclusion/exclusion criteria to get more reliable and generalized outcomes for preoperative clinical decisions in HCC patients.

ACKNOWLEDGEMENTS

The author acknowledges Central University of Punjab, Bathinda, Punjab, India for providing necessary infrastructure.