The global incidence of melanoma has been increasing gradually, presenting a considerable public health challenge (1, 2). The incidence rate of melanoma vary geographically, with the highest rates observed in Australia and New Zealand, followed by North America and Europe (3). Notably, in Asian populations, although the overall incidence remains lower, melanoma frequently manifests as acral or mucosal subtypes, which are associated with poorer prognosis (4).

Surgical excision is the most prevalent treatment modality for melanoma. However, postoperative recurrence rates remain substantial (5). This high recurrence risk necessitates the exploration of effective adjuvant therapies to improve long-term outcomes. The challenge is particularly pronounced in Chinese populations, where acral and mucosal melanoma subtypes are more prevalent (6, 7). These subtypes exhibit lower response rates to immune checkpoint inhibitors like anti-PD-1 and anti-CTLA4 antibodies (7, 8), resulting in inferior clinical prognosis for patients with advanced melanoma. Consequently, there is a pressing need to develop alternative therapeutic method that can improve clinical outcomes, particularly for melanoma in Chinese populations.

Among the adjuvant therapeutic strategies, the interferon-α1b (IFN-α1b) has revealed promising treatment effect. As a member of the interferon α family, IFN-α1b exhibits relatively favorable tolerability, eliciting anti-tumor responses and potentially playing a central role in innate immunity, thereby providing a rationale for its clinical application in oncology (9–12). Recently, our team has demonstrated the promising potential of human IFN-α1b in melanoma treatment. A retrospective study by Shi et al. showed that IFN-α1b monotherapy exhibited favorable outcomes in unresectable stage IV melanoma patients, resulting in a median overall survival (mOS) of 14.1 months (13). Subsequently, Gao et al. reported the efficacy of adjuvant IFN-α1b in resected stage IIIB or IIIC melanoma. This study revealed promising recurrence-free survival rates of 75.4%, 47.4%, and 37.2% at 12, 24, and 36 months, respectively. Moreover, the overall survival rates were impressive, with 100%, 81.9%, and 71.5% at the same time points (14). Despite these encouraging results, several important questions remain. The optimal criteria for patient selection and the long-term efficacy of IFN-α1b as an adjuvant therapy remain to be fully elucidated. Furthermore, given the heterogeneity of melanoma and the variability of treatment responses, there is a pressing need for more sophisticated prognostic tools to guide treatment decisions and predict outcomes in patients receiving adjuvant IFN-α1b therapy.

In recent years, machine learning has emerged as a promising tool in oncology, providing novel opportunities for predicting treatment outcomes and personalizing patient care (15). By analyzing intricate patterns within large datasets, machine learning algorithms have the potential to identify subtle prognostic factors and treatment response indicators that may not be discernible through traditional statistical methods (16). This approach is particularly promising in the context of melanoma, where the heterogeneity of the disease and the variability in treatment responses present substantial challenges to clinical decision-making (17). Based on these considerations, our study aims to evaluate the efficacy of postoperative IFN-α1b in melanoma patients. More importantly, we want to develop and validate machine learning models for assessing prognosis of these patients, by integrating clinical data with advanced machine learning methodologies, so as to optimize melanoma management and improve the outcome of patients with melanoma, especially in Chinese populations.

In this retrospective study, we demonstrated the efficacy of adjuvant IFN-α1b therapy in improving clinical outcomes for patients with resected stage III-IV melanoma. Our analysis revealed a median RFS of 20 months and a median OS of 81 months, underscoring the potential of IFN-α1b in the management of advanced melanoma. To further refine our understanding and prediction of treatment efficacy, we developed and compared five machine learning models. Among these, the DT model demonstrated superior performance in predicting individual patient responses to IFN-α1b therapy. This innovative approach not only validated the overall efficacy of the treatment but also provided a novel tool for personalized treatment decisions in melanoma management.

The efficacy of IFN-α1b in melanoma treatment can be attributed to its multifaceted biological effects. Primarily, IFN-α1b exerts direct antiproliferative effects on tumor cells by inducing cell cycle arrest and apoptosis (18). Additionally, it enhances the immune response against melanoma cells through multiple mechanisms: activating natural killer cells, upregulating the expression of major histocompatibility complex (MHC) class I molecules, and promoting the differentiation of dendritic cells (19–21). Recent studies have also highlighted the role of IFN-α1b in modulating the tumor microenvironment, including the inhibition of angiogenesis and the promotion of a more immunogenic tumor phenotype (18, 19). These mechanisms contribute to the observed clinical benefits in our study, reinforcing the rationale for IFN-α1b as an adjuvant therapy in advanced melanoma.

Our findings align with and extend previous studies on interferon therapy in melanom. A meta-analysis by Ives et al. demonstrated a significant improvement in RFS with interferon therapy, consistent with our results (22). However, our study uniquely contributes to the field by employing machine learning techniques to predict individual patient responses, an approach not previously applied in this context. While recent research has focused on alternative immunotherapies like PD-1 inhibitors (23), our study reaffirms the value of IFN-α1b, particularly in melanoma patients in Chinese population who may exhibit lower response rates to immune checkpoint inhibitors. Moreover, our machine learning approach offers a novel strategy for patient stratification, potentially optimizing the use of IFN-α1b and may address the ongoing challenge of treatment selection in advanced melanoma.

The application of machine learning, particularly the DT model, represents a significant methodological advancement in predicting IFN-α1b treatment efficacy. The DT model offers several advantages over traditional statistical methods in this context. Firstly, they can capture non-linear relationships and complex interactions between variables, which are often present in biological systems (24). Secondly, the DT model provides easily interpretable results, allowing clinicians to understand the decision-making process, a crucial factor in medical applications (25). The hierarchical structure of the DT model also aligns well with clinical decision-making processes, making the model’s predictions more intuitive (26). Furthermore, DTs are robust to outliers and missing data, common challenges in clinical datasets (27). In our study, the DT model outperformed traditional logistic regression in both accuracy and area under the receiver operating characteristic curve (AUC-ROC), demonstrating its superior predictive capability in this complex clinical scenario.

Our machine learning model identified serum albumin levels as the most significant predictor of treatment response and prognosis in melanoma patients receiving IFN-α1b therapy. This finding aligns with emerging evidence on the crucial role of albumin in cancer biology and treatment outcomes. Albumin serves as a key indicator of nutritional status and overall health, factors known to influence cancer prognosis (28). A large-scale study by Gupta et al. found that pretreatment serum albumin levels were independently associated with overall survival in cancer patients across multiple tumor types (28). In the study conducted by Xie et al., the overall survival was significantly higher in the group with elevated albumin levels than in the lower albumin levels (29). Moreover, recent studies have elucidated albumin’s direct effects on tumor biology. Serum albumin has been shown to modulate the tumor microenvironment by influencing oxidative stress and inflammatory responses (30). It also plays a role in drug transport and metabolism, potentially affecting the pharmacokinetics and efficacy of IFN-α1b (31). Overall, serum albumin levels may serve as a biomarker reflecting the inflammatory status and nutritional state of patients. Low albumin levels could indicate a heightened inflammatory response or malnutrition, both of which are known to adversely affect melanoma prognosis. These multifaceted functions of albumin underscore its importance as a predictive biomarker in our model and suggest potential avenues for therapeutic interventions aimed at optimizing treatment outcomes in melanoma patients.

Despite the promising results, our study has several limitations that warrant consideration. Firstly, the retrospective, single-center design and non-random selection of patients introduce potential biases and limit the generalizability of our findings. Selection bias may have influenced patient inclusion, potentially affecting the observed treatment outcomes. The completeness and accuracy of clinical records may introduce errors that could affect the validity of our findings. What’s more, the single-center nature of the study also raises questions about the applicability of our machine learning model to diverse patient populations and clinical settings (32). Additionally, while our model showed good predictive performance, the relatively small sample size may have limited its ability to capture rarer prognostic factors or subgroup effects. The DT model demonstrated superior performance, likely due to the dataset’s characteristics, which may be more conducive to simpler models. Complex models, while potentially more powerful, risk overfitting, particularly in datasets with limited sample sizes. Future studies should explore the balance between model complexity and generalizability. Lastly, the evolving landscape of melanoma treatment, with the introduction of targeted therapies and novel immunotherapies, may impact the long-term relevance of our findings focused on IFN-α1b. Future multi-center, prospective studies with larger cohorts are needed to validate and refine our predictive model.

In conclusion, our study demonstrates the efficacy of IFN-α1b adjuvant therapy in melanoma and develops the DT model, which offers a promising tool for personalized risk assessment. The identification of serum albumin as a key predictive factor offers new insights into the biological mechanisms underlying treatment efficacy.