Risk Prediction Based on Clinicopathologic Features in Korean Melanoma Patients by Machine Learning

Taeho Yuh; Hyunwook Kim; Eun Sil Baek; Joong Bae Ahn; Dongwoo Chae; Kee Yang Chung; Byungho Oh; Mi Ryung Roh; Sang Joon Shin

doi:10.4143/crt.2025.1154

Articles

Page Path: HOME > J Korean Cancer Assoc > Ahead-of print articles > Article

Original Article Risk Prediction Based on Clinicopathologic Features in Korean Melanoma Patients by Machine Learning: Taeho Yuh¹, Hyunwook Kim^1,2, Eun Sil Baek³, Joong Bae Ahn^1,2,3, Dongwoo Chae⁴, Kee Yang Chung⁵, Byungho Oh⁵, Mi Ryung Roh⁵, Sang Joon Shin^1,2,3; DOI: https://doi.org/10.4143/crt.2025.1154
Published online: May 7, 2026

¹Department of Internal Medicine, Yonsei University College of Medicine, Seoul, Korea

²Division of Medical Oncology, Yonsei Cancer Center, Seoul, Korea

³Sondang Institute for Cancer Research, Yonsei University College of Medicine, Seoul, Korea

⁴Department of Pharmacology, Yonsei University College of Medicine, Seoul, Korea

⁵Department of Dermatology and Cutaneous Biology Research Institute, Yonsei University College of Medicine, Seoul, Korea

Correspondence: Sang Joon Shin, Division of Medical Oncology, Department of Internal Medicine, Yonsei Cancer Center, Yonsei University College of Medicine, 50-1 Yonsei-ro, Seodaemun-gu, Seoul 03722, Korea
Tel: 82-2-2228-8053 E-mail: ssj338@yuhs.ac

• Received: October 22, 2025 • Accepted: May 4, 2026

This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/4.0/) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.

791 Views
40 Download

Full Article

Download PDF

Abstract

Purpose
The prognosis of malignant melanoma differs across ethnic groups, partly due to variations in histologic subtypes. Accurate risk stratification at diagnosis is crucial but remains challenging in Korea due to its low incidence. This study aimed to develop and evaluate machine learning (ML) models for predicting 2-year disease-free survival (DFS) using clinicopathologic features available at the time of diagnosis.
Materials and Methods
A retrospective cohort of 1,657 Korean melanoma patients (2006-2023) was analyzed. Clinical and pathological variables—excluding TNM staging and treatment data—were used to train ML models, including decision tree, random forest, boosting methods, XGBoost, and deep neural networks. Performance was evaluated using accuracy, precision, recall, F1 score, and AUC.
Results
The study included 1,657 patients with 12 clinicopathologic variables, showing a median DFS of 63.9 months (95% confidence interval [CI], 52.93 to 85.33) and a 2-year survival rate of 67.65% (95% CI, 65.16 to 70.00). Among the evaluated algorithms, XGBoost demonstrated the highest F1 score of 0.761. Feature importance analysis identified primary tumor site, mitotic rate, and body mass index (BMI) as the most influential predictors of survival. Kaplan-Meier curves and Cox proportional hazards analysis confirmed the clinical relevance of these features.
Conclusion
The XGBoost model showed the best performance for predicting 2-year DFS probability in Korean melanoma patients. Primary cancer site, mitotic rate, and BMI were identified as significant features influencing survival.
Key words: Malignant melanoma, Machine learning, Disease-free survival, Korean population, Prognostic model

Introduction

Malignant melanoma is the most rapidly progressing type of skin cancer and is the fifth most common cancer in the United States [1]. In South Korea, 673 new cases were diagnosed in 2024 alone [2]. Early and accurate diagnosis and risk stratification are essential due to its aggressive clinical course [3]. Surgical excision remains the cornerstone for histopathological diagnosis and local control, typically accompanied by sentinel lymph node biopsy for staging.

Current clinical guidelines rely on the tumor-node-metastasis (TNM) staging system, particularly emphasizing features such as metastasis, tumor thickness, ulceration, and lymph node involvement [4,5]. However, because these staging components are such strong determinants of prognosis, they may mask the potential contribution of other clinically informative variables.

Importantly, malignant melanoma exhibits marked heterogeneity across ethnic groups in terms of pathogenesis, subtype distribution, and clinical outcomes [6]. In Asian populations, particularly South Koreans, acral lentiginous melanoma (ALM)—unrelated to ultraviolet radiation and often arising on palms, soles, and nail beds—predominates [7]. ALM is frequently diagnosed at advanced stages and shows higher rates of ulceration, nodal involvement, and metastasis, contributing to worse outcomes compared to superficial spreading melanoma (SSM) common in Caucasians [8-10].

Melanoma is considered rare in Korea, limiting access to large-scale datasets and hindering robust, population-specific prognostic models. Machine learning (ML) methods provide powerful tools for integrating heterogeneous clinical variables, capturing non-linear relationships, and identifying complex prognostic patterns that may not be apparent in traditional statistical models.

Although prior studies have applied ML to various melanoma-related tasks—including image-based lesion classification, mutation prediction, and survival estimation— few have focused specifically on risk stratification using structured clinical data in non-Western populations [11-16]. Moreover, many existing models rely heavily on TNM staging, which, while powerful, may overshadow other clinically meaningful variables that are available at the time of diagnosis. Although TNM staging remains the standard-of-care prognostic framework in melanoma, we aimed to investigate whether clinicopathologic features available at diagnosis could provide complementary stage-independent information for predicting 2-year disease-free survival (DFS). Thus, the proposed model is intended as an adjunctive exploratory tool rather than a replacement for TNM-based prognostication.

In this study, we aimed to develop and validate an ML-based model to predict 2-year DFS in South Korean patients with stage I-III malignant melanoma, using exclusively clinicopathologic features independent of TNM staging. Importantly, we deliberately excluded TNM staging information to evaluate the prognostic value of non-staging variables in early risk stratification. By analyzing model performance and feature importance, we identified both tumor-intrinsic (e.g., mitotic rate, histologic subtype) and host-related (e.g., body mass index [BMI] and C-reactive protein [CRP]) predictors that contribute to DFS.

Unlike most previous prognostic models, which rely heavily on TNM staging or Western patient cohorts, our work demonstrates the feasibility of TNM-free, population-specific risk modeling in Asian cohorts. These findings may contribute to the development of more population-specific and clinically accessible prognostic tools for melanoma patients at the time of diagnosis.

Materials and Methods

1. Design and participants

Medical records of patients with histologically confirmed malignant melanoma who received care at Severance Hospital, South Korea, between January 1, 2006, and October 31, 2023, were collected. A total of 1,910 patients aged 19 years or older were included. Data were extracted from electronic health records (EHRs) using text mining and natural language processing techniques. The dataset included detailed information on demographics, surgical history, pathology reports, laboratory test results, treatment records, and clinical outcomes.

All data were anonymized and securely stored on an encrypted virtual server. Variables collected included: patient ID (anonymized), date of birth, sex, date of histological diagnosis, age at diagnosis, primary tumor site, baseline BMI, baseline hematologic parameters (hemoglobin [Hb], neutrophil-to-lymphocyte ratio [NLR], platelet count [PLT], lactate dehydrogenase [LDH], and CRP), histological classification, TNM staging, Breslow thickness, ulceration, mitotic rate, metastasis status, molecular profiles (BRAF, NRAS, and programmed cell death ligand 1 [PD-L1]), surgical information (date, type, and lymph node dissection), recurrence data, use of radiotherapy, adjuvant and palliative chemotherapy, disease progression status, and survival outcomes.

For binary 2-year DFS classification, patients who experienced recurrence or death within 2 years of diagnosis were labeled as events, whereas patients who remained event-free for at least 2 years were labeled as non-events. Patients with less than 2 years of follow-up and no documented event were excluded because definitive 2-year DFS status could not be assigned.

2. Endpoints

The primary endpoint of this study was to develop an ML-based risk stratification model capable of predicting 2-year DFS probability in stage I-III patients diagnosed as malignant melanoma. Model performance was evaluated and validated using a held-out test and validation set.

The secondary endpoint was to assess the clinical relevance of input features through two approaches: (1) analyzing feature importance derived from the best-performing ML model, and (2) conducting univariate and multivariate Cox proportional hazards (CPH) analyses to determine the statistical association between individual features and survival outcomes.

These findings were compared against established oncologic knowledge and real-world patient characteristics to evaluate their interpretability and potential clinical utility.

3. Machine learning

All ML models were implemented using Python 3.12 and Jupyter Notebook. Six algorithms were applied: decision tree, random forest, bagging, AdaBoost, GradientBoost, and XGBoost [17]. To avoid model bias from direct treatment effects, variables such as surgery details, recurrence, radiotherapy, and chemotherapy were excluded. Furthermore, variables either sparsely available (e.g., NRAS and PD-L1) or overly dominant in survival prediction (e.g., pathologic TNM staging, especially M stage) were excluded to reduce bias and enhance interpretability of feature importance.

Categorical variables were encoded using label encoding. Missing values were handled by imputing numerical variables using median values, while categorical variables with missing entries were assigned a separate “unknown” category. The extent of missingness for each variable is provided in S1 Table. The dataset was randomly divided into training, validation, and test sets in an 8:1:1 (n=1,325, n=166, n=166) ratio. The training set was used to fit the models, while the validation set was utilized exclusively for hyperparameter tuning and model selection. The test set was held out throughout the training process to provide an unbiased estimate of model performance. Model performance was evaluated using five standard classification metrics: area under the receiver operating characteristic curve (ROC AUC), accuracy, precision, recall, and F1 score.

For tree-based models, key hyperparameters were optimized using grid search with 5-fold cross-validation. The exact hyperparameter grids searched are provided in S2 Table. All models were tuned using grid search approaches optimized for F1 score. All models were trained and evaluated using scikit-learn (v1.3) and XGBoost (v1.7) libraries in Python 3.12.

In addition to these classical models, a deep neural network (DNN) was implemented using the Keras library with a TensorFlow backend. Hyperparameter optimization was conducted using grid search. The DNN model was trained using binary cross-entropy loss and early stopping based on validation loss. The best performance was achieved with batch_size=16 and epochs=100.

4. Statistical analysis

Kaplan-Meier survival curves were constructed to assess survival differences by sex, primary tumor site, and histological subtype. For continuous and ordinal variables, univariate and multivariate CPH models were fitted to estimate hazard ratios. Variables tested included: sex, age at diagnosis, BMI, Eastern Cooperative Oncology Group (ECOG) performance status, Hb, NLR, PLT, LDH, CRP, mitotic rate, primary cancer site, and histological classification. A two-sided p-value of < 0.05 was considered statistically significant.

Results

1. Patient characteristics

Between January 1, 2006, and October 31, 2023, data were collected from 1,910 South Korean patients with histologically confirmed malignant melanoma. To eliminate the influences from other malignancies, 120 patients were excluded due to: post-melanoma diagnoses of other cancers within the study period (n=98), prior diagnoses of stage IV cancers any period before the study (n=6), treatment history for subsequently diagnosed cancers within the study period (n=11), recent cancer treatment within 1 year (n=4), or cancer diagnosis within the past year (n=1). Additionally, 133 stage IV melanomas at diagnosis were also excluded to focus the analysis on early- to intermediate-stage disease. After exclusions, a total of 1,657 patients were included for analysis and model training (S3 Fig.).

Baseline characteristics are summarized in Table 1. The median age was 59 years, ranging from 19 to 97 years, with female predominance (53.0%). The primary tumor site was cutaneous in 1,015 patients (61.3%), ocular in 429 (25.9%), and mucosal in 213 (12.9%). Among cutaneous melanomas, histological subtypes included acral lentiginous (22.0%), superficial spreading (9.7%), nodular (6.3%), and lentigo maligna (0.9%). ECOG performance status was available for 940 patients (56.7%), with the majority classified as ECOG 0 (841, 50.8%). Pathological staging data were available for 1,182 patients (71.3%), with stage II (27.0%) being the most common, followed by stage I (19.4%), stage III (14.2%), and carcinoma in situ (10.7%). Based on the World Health Organization (WHO) BMI classification, normal weight (18.5-24.9 kg/m²) was observed in 939 patients (56.7%). The remaining patients were classified as follows: underweight (< 18.5 kg/m²), 45 patients (2.7%); overweight (25.0-29.9 kg/m²), 362 patients (21.9%); and obese (≥ 30.0 kg/m²), 311 patients (18.8%). At data cutoff, 555 patients (33.5%) had experienced relapse, and 559 patients (33.7%) had expired.

The total median DFS was 63.9 months (95% confidence interval [CI], 52.9 to 85.3), with a 2-year survival rate of 67.7% (95% CI, 65.2 to 70.0). The median follow-up duration was 37.3 months (95% CI, 34.6 to 38.9) (Kaplan-Meier survival estimates are shown in S4 Fig.).

2. ML model performance

Data from 1,657 patients were used to develop models predicting 2-year DFS using seven supervised learning algorithms: decision tree, random forest, bagging, AdaBoost, GradientBoost, XGBoost, and a DNN implemented via Keras. The input variables included demographic, laboratory, and pathological features (e.g., sex, age at diagnosis, BMI, ECOG, Hb, NLR, PLT, LDH, CRP, tumor site, histological subtype, and mitotic rate).

For the DNN, a feedforward architecture was used consisting of two hidden layers with ReLU activation, dropout regularization (rate=0.5), and a sigmoid output layer. Hyperparameters were optimized via grid search, resulting in a batch size of 16, 100 epochs, and the Adam optimizer.

Model performance was assessed using a randomized test set and evaluated across five metrics: ROC AUC, accuracy, precision, recall, and F1 score. As shown in Table 2, the XGBoost model outperformed the other algorithms, achieving the highest recall (0.914) and F1 score (0.761). The DNN model yielded an ROC AUC of 0.605 and an F1 score of 0.724.

Given the moderate class imbalance in DFS outcomes, the F1 score was considered the most appropriate performance metric [18], suggesting that XGBoost is the optimal model for predicting 2-year DFS in stage I-III Korean patients with newly diagnosed melanoma.

Taken together, these results demonstrate that XGBoost outperformed other models in predicting 2-year DFS in Korean melanoma patients, achieving an F1 score of 0.761. This model, which excludes TNM staging and relies solely on clinicopathologic features available at diagnosis, highlights the potential of ML models to support early risk stratification in melanoma patients. The strong performance suggests that a TNM-independent model can still yield clinically meaningful predictions.

3. Feature importance

To interpret model predictions and identify key prognostic factors, feature importance was evaluated across all models (Fig. 1). In the best-performing XGBoost model, the most influential predictors included primary site, mitotic rate, and baseline BMI. Other models showed consistent trends, with baseline NLR gaining more importance in decision tree and random forest models. These findings highlight the relevance of both tumor-intrinsic and host-related features in predicting survival outcomes.

Similarly, to interpret the DNN model, we applied SHAP (SHapley Additive exPlanations) using the `DeepExplainer` module from the SHAP Python library. The SHAP values were computed for each input feature, allowing estimation of the marginal contribution of each feature to the model’s prediction of 2-year DFS. The top contributing features were visualized using SHAP summary plots (S5 Fig.).

4. Survival analysis based on key features

To validate the prognostic relevance of variables identified through ML, univariate CPH regression was performed (Table 3). Given the non-linear relationship between BMI and survival, analyses were conducted using WHO BMI classification categories: underweight (< 18.5 kg/m²), normal (18.5-24.9), overweight (25.0-29.9), and obese (≥ 30.0). The following variables were significantly associated with reduced survival: sex, age at diagnosis, LDH, NLR (all p < 0.005), and CRP (p=0.005). ECOG performance score (p=0.447), Hb (p=0.412) and PLT (p=0.139) did not reach statistical significance in this model.

Univariate CPH showed that mucosal melanoma (hazard ratio [HR], 2.564; p < 0.001) and nodular melanoma (HR, 1.700; p < 0.001) were associated with significantly worse survival compared to the reference, ALM. In contrast, ocular melanoma was associated with improved survival (HR, 0.442; p < 0.001). SSM (HR, 0.846; p=0.256) and lentigo maligna (HR, 0.365; p=0.155) did not show significant differences in survival relative to the reference.

The relationship between BMI, using WHO classification, and DFS followed a J-shaped curve. Patients with BMI in the range of 25.0-29.9 kg/m² (overweight) had the most favorable survival (HR, 0.694; p < 0.001), while those classified as obese (BMI ≥ 30) showed significantly poorer survival (HR, 2.323; p < 0.001). The underweight patient group (BMI < 18.5) did not show a significant difference in survival compared to the normal group (HR, 0.993; p=0.973).

Kaplan-Meier curves of categorical features, along with BMI classification and pathological staging, were analyzed to visualize the survival comparison (Fig. 2). Median DFS, 2-year DFS rate, and log-rank test results are presented in Table 4.

Discussion

This study developed and validated an ML-based model for predicting 2-year DFS in stage I-III South Korean patients with malignant melanoma using structured clinical data available at diagnosis. Using a structured EHR-derived dataset from a large, single-center cohort, a total of seven supervised learning models, including classical tree-based algorithms and a DNN, were implemented, and the best-performing model was selected based on F1 score, which is appropriate given the moderate class imbalance in survival outcomes. Among the models tested, XGBoost achieved the most balanced performance across multiple metrics, including recall (0.914) and F1 score (0.761). Feature importance analysis from the XGBoost model revealed several key predictors of survival, notably primary site, mitotic rate, and baseline BMI.

Previous studies have explored the utility of ML in melanoma for various tasks, including lesion classification from dermoscopic images, mutation prediction, and treatment response modeling [14,19,20]. However, fewer studies have focused specifically on risk stratification using structured clinical variables, particularly outside of Western populations.

One multicenter study by Cozzolino et al. [21] applied a range of ML models including random forest, support vector machine, and DNN to a U.S. dataset. They presented DNN (balanced accuracy 91.1%, F1 score 80.0%) and random forest (balanced accuracy 88.0%, F1 score 59.6%) as the models with the highest performance. Age, mitotic rate, T4 staging, presence of ulceration, and metastasis were highlighted as feature importances in random forest. Compared to the study by Cozzolino et al. [21], in which a DNN achieved the highest performance (F1 score 80.0%, balanced accuracy 91.1%) among several models, our XGBoost-based approach showed a relatively lower F1 score. However, it is noteworthy that the key predictors in this model included TNM-related variables such as T4 staging and metastasis, and their dataset encompassed stage IV patients. In contrast, our study focused exclusively on stage I-III patients and intentionally excluded all TNM staging features to explore DFS prediction from a different perspective. This distinction may explain the performance gap, but also highlights the strength of our approach in uncovering alternative, clinically relevant predictors in earlier-stage disease.

Another study by Wan et al. [22] shares methodological similarities with ours in applying ML to structured clinical and histopathologic data. Their gradient boosting model achieved strong ROC AUC values (0.845 internal, 0.812 external), with Breslow thickness, mitotic rate, and age as top predictors. However, their analysis was restricted to cutaneous melanoma. In contrast, our study incorporated cutaneous, ocular, and mucosal subtypes, and identified primary site as one of the most important features both in ML models and CPH analysis. This broader scope adds value by demonstrating how anatomical subtype contributes meaningfully to DFS prediction across diverse melanoma presentations. Furthermore, our study also highlights the nonlinear correlation of baseline BMI and survival, showing deeper insight into baseline patient characteristics.

To contextualize our model performance, we compared our results with prior Korean and international studies that utilized either TNM-based models or ML classifiers incorporating staging variables. Despite excluding TNM variables, our XGBoost model achieved a competitive F1 score of 0.761, which was close to or exceeded the performance of existing models (Table 5). Comparisons with prior studies should be interpreted cautiously because of substantial differences in study population, stage distribution, endpoints, and performance metrics.

Feature importance analysis from the XGBoost model revealed several key predictors of survival, notably primary site, mitotic rate, and baseline BMI. These findings diverge from previous studies that emphasized TNM-associated factors such as metastasis, ulceration, or tumor thickness—variables intentionally excluded in our study to avoid circular reasoning. By omitting TNM staging components, which are already well-established determinants of prognosis and used for clinical decision-making, we aimed to develop a model that could reveal complementary, staging-independent risk signals from baseline clinicopathologic data. These features were also statistically significant in CPH analysis, reinforcing their independent prognostic value.

In comparing melanoma subtypes, mucosal melanoma is known to carry the poorest prognosis [23,24], whereas ocular melanoma has been reported to show higher 5-year recurrence-free survival (RFS) and overall survival (OS) rates compared to cutaneous melanoma (RFS: 56.8% vs. 30.1%; OS: 76.6% vs. 68.8%) [25]. These findings align with our results, wherein mucosal subtype was associated with significantly worse survival compared to acral lentiginous and ocular subtypes, reinforcing the heterogeneity in prognosis based on anatomical origin. Notably, among the subtypes of cutaneous melanoma, acral lentiginous and nodular have been reported to show poorer outcomes compared to superficial spreading [26-28]. This partially aligns with our study, in which nodular melanoma showed significantly worse outcomes, whereas acral lentiginous and superficial spreading showed no difference.

The prognostic significance of mitotic rate has been wellestablished: a large American Joint Committee on Cancer melanoma cohort study (n=13,296, stages I-II) reported that mitotic rate was the second strongest independent predictor of melanoma-specific survival after tumor thickness. Patients with ≥ 20 mitoses/mm2 had a 10-year survival of only 48% compared to 93% in those with 0 mitoses/mm2 (p< 0.001) [29].

The role of BMI in melanoma prognosis has been debated. Earlier reviews suggested mixed associations between obesity and disease-free interval or melanoma-specific survival [30]. More recent studies, particularly involving immune checkpoint inhibitor–treated patients, report an “obesity paradox”—where higher BMI correlates with better treatment response and survival [31]. In our study, we observed a J-shaped relationship between BMI and DFS: overweight patients exhibited better outcomes, while obese patients had worse survival compared to those with normal BMI. This pattern may reflect the complex, dual role of adiposity in cancer biology—balancing potential metabolic or immunologic advantages at moderate levels with detrimental effects at higher extremes. Obesity has been implicated in impaired immune surveillance, altered cytokine profiles (e.g., interleukin-6 and tumor necrosis factor α), and reduced therapy responsiveness. Additionally, residual confounding factors such as performance status, comorbidities, and treatment variations—unmeasured in our dataset—may partially account for this non-linear trend. Further research in larger, multi-institutional cohorts with more comprehensive clinical data is needed to validate and clarify these findings.

While we adopted established algorithms such as XGBoost and DNN, this choice was deliberate. For structured, tabular datasets with moderate sample sizes—as in our cohort of 1,657 patients—tree-based ensemble models such as XGBoost have consistently outperformed more complex deep learning or hybrid architectures in both accuracy and interpretability [32]. Furthermore, the clinical context requires transparent models where feature contribution is readily interpretable for clinicians; this made tree-based models a strong candidate due to their well-characterized feature importance metrics. In contrast, recent novel or hybrid architectures such as TabNet, TabTransformer, or attention-based models offer potential gains in performance but often require much larger datasets and involve greater risk of overfitting in low-dimensional clinical settings. Moreover, they lack the maturity and interpretability tooling that tree-based models currently offer. While our current study focused on diagnostic-time risk stratification using readily available features, we acknowledge that future studies involving multi-institutional or multimodal data may benefit from benchmarking against such advanced methods.

Several limitations should be acknowledged. First, this study was based on data from a single tertiary referral center in Korea, which may limit generalizability. Second, genomic variables such as BRAF, NRAS, and PD-L1 mutation status were excluded due to data sparsity, and treatment-related variables were intentionally omitted to focus on baseline prognostication. Third, the absence of external validation restricts conclusions about real-world model applicability. Despite internal cross-validation, future studies should apply these models to independent external cohorts to evaluate generalizability. Additionally, deeper integration of multi-modal data—including imaging, longitudinal lab values, and richer genomic profiling—may further improve model performance and risk classification.

Because DFS is inherently a time-to-event outcome, reducing it to a binary 2-year endpoint does not fully account for censoring. Excluding patients with insufficient follow-up but no event may have introduced selection bias or informative censoring. Future studies using survival-specific modeling approaches may provide a more comprehensive assessment.

Patients with other malignancies were excluded to reduce confounding from unrelated cancer treatments, surveillance patterns, and competing oncologic events, thereby enabling a more homogeneous analysis of stage I-III melanoma outcomes. Nevertheless, this exclusion may have introduced selection bias toward a healthier cohort.

Nonetheless, this study demonstrates that ML can be effectively applied to structured clinical data to develop a high-performing, interpretable prognostic model for malignant melanoma in Asian populations. Unlike prior studies that relied on image-based or Western-biased datasets, our analysis leveraged population-specific variables and validated the clinical relevance of key predictors through both ML and conventional survival analysis. The exclusion of TNM-related features allowed the model to focus on less-explored but clinically available variables such as inflammatory markers, hematologic indices, and histologic subtype. This approach increases the interpretability of the model, avoids overfitting to known staging elements, and may facilitate early risk estimation prior to the completion of full staging workup. These findings highlight the potential of combining ML with EHR data to enhance risk stratification and guide early treatment planning in rare cancer settings.

In conclusion, this study demonstrates the feasibility of applying ML to structured EHR data for risk stratification in melanoma patients. The XGBoost model showed the best predictive performance among the evaluated algorithms and identified several clinically meaningful predictors, including primary tumor site, mitotic rate, and baseline BMI. Although TNM staging remains the cornerstone of melanoma prognostication, our TNM-independent framework highlights additional clinicopathologic signals that may contribute to early risk assessment. Future studies incorporating external validation and survival-specific modeling approaches are warranted to further improve predictive performance and generalizability.

Electronic Supplementary Material

Supplementary materials are available at Cancer Research and Treatment website (https://www.e-crt.org).

CRT_25-1154_S1_Table.pdf

CRT_25-1154_S2_Table.pdf

CRT_25-1154_S3_Fig.pdf

CRT_25-1154_S4_Fig.pdf

CRT_25-1154_S5_Fig.pdf

NOTES

Ethical Statement

This study was conducted in accordance with the principles outlined in the Declaration of Helsinki and was approved by the Institutional Review Board (IRB) of Severance Hospital, Yonsei University Health System (IRB No. 4-2023-1504; Jan 3, 2024). Given the retrospective nature of the study and the use of fully anonymized data, informed consent was waived by the IRB. All data handling and analyses complied with institutional data protection policies and national regulations regarding patient confidentiality and privacy.

Author Contributions

Conceived and designed the analysis: Yuh T, Shin SJ.

Collected the data: Yuh T, Baek ES.

Contributed data or analysis tools: Yuh T.

Performed the analysis: Yuh T.

Wrote the paper: Yuh T.

Review & editing: Kim H, Ahn JB, Chae D, Chung KY, Oh B, Roh MR, Shin SJ.

Conflicts of Interest

Conflict of interest relevant to this article was not reported.

Fig. 1.

Feature importances of decision tree (A), random forest (B), AdaBoost (C), GradientBoost (D), and XGBoost (E), computed from model-specific metrics such as Gini impurity or information gain. BMI, body mass index; CRP, C-reactive protein; ECOG, Eastern Cooperative Oncology Group; Hb, hemoglobin; LDH, lactate dehydrogenase; NLR, neutrophil-to-lymphocyte ratio; PLT, platelet count.

Fig. 2.

Kaplan-Meier curves of patients compared by sex (A), pathological stage (B), primary site (C), histological type (D), and baseline body mass index (BMI) (E). All survival curves were estimated using the ‘lifelines’ package, and significance was evaluated via log-rank test.

Table 1.

Baseline demographic and clinicopathologic characteristics of the study cohort

	Total	Cutaneous					Ocular	Mucosal
	Total	Total	Acral lentiginous	Superficial spreading	Nodular	Lentigo maligna	Ocular	Mucosal
Total	1,657	1,015 (61.3)	365 (22.0)	160 (9.7)	104 (6.3)	15 (0.9)	429 (25.9)	213 (12.9)
Age (yr)
Median (range)	59 (19-97)	60 (19-95)	63 (21-93)	58 (19-87)	58 (24-87)	70 (42-86)	56 (19-90)	62 (28-97)
Sex
Male	778 (47.0)	469 (46.2)	171 (46.8)	70 (43.8)	52 (50.0)	10 (66.7)	206 (48.0)	103 (48.4)
Female	879 (53.0)	546 (53.8)	194 (53.2)	90 (56.3)	52 (50.0)	5 (33.3)	223 (52.0)	110 (51.6)
ECOG PS
0	841 (50.8)	575 (56.7)	254 (69.6)	121 (75.6)	68 (65.4)	12 (80.0)	173 (40.3)	93 (43.7)
1	90 (5.4)	18 (1.8)	3 (0.8)	2 (1.3)	1 (1.0)	0	54 (12.6)	18 (8.5)
2	7 (0.4)	4 (0.4)	1 (0.3)	0	0	0	1 (0.2)	2 (0.9)
3	2 (0.1)	1 (0.1)	0	0	0	0	0	1 (0.5)
Pathological stage
In situ	177 (10.7)	174 (17.1)	96 (26.3)	17 (10.6)	0	6 (40.0)	0	3 (1.4)
I	321 (19.4)	242 (23.8)	102 (28.0)	66 (41.3)	13 (12.5)	6 (40.0)	66 (15.4)	13 (6.1)
II	448 (27.0)	274 (27.0)	111 (30.4)	43 (26.9)	47 (45.2)	1 (6.7)	153 (35.7)	21 (9.9)
III	236 (14.2)	177 (17.4)	48 (13.2)	29 (18.1)	40 (38.5)	0	30 (7.0)	29 (13.6)
Baseline BMI
Underweight	45 (2.7)	25 (2.5)	6 (1.6)	5 (3.1)	2 (1.9)	0	15 (3.5)	5 (2.3)
Normal	939 (56.7)	554 (54.6)	217 (59.5)	109 (68.1)	50 (48.1)	12 (80.0)	279 (65.0)	106 (49.8)
Overweight	362 (21.9)	209 (20.6)	92 (25.2)	30 (18.8)	26 (25.0)	2 (13.3)	104 (24.2)	49 (23.0)
Obese	311 (18.8)	227 (22.4)	50 (13.7)	16 (10.0)	26 (25.0)	1 (6.7)	31 (7.2)	53 (24.9)
Relapse
No	1,102 (66.5)	658 (64.8)	270 (74.0)	116 (72.5)	46 (44.2)	13 (86.7)	334 (77.9)	110 (51.6)
Yes	555 (33.5)	357 (35.2)	95 (26.0)	44 (27.5)	58 (55.8)	2 (13.3)	95 (22.1)	103 (48.4)
Death
No	1,098 (66.3)	689 (67.9)	293 (80.3)	130 (81.3)	66 (63.5)	14 (93.3)	315 (73.4)	94 (44.1)
Yes	559 (33.7)	326 (32.1)	72 (19.3)	30 (18.8)	38 (36.5)	1 (6.7)	114 (26.6)	119 (55.9)

Values are presented as number (%) unless otherwise indicated. BMI, body mass index; ECOG PS, Eastern Cooperative Oncology Group performance status.

Table 2.

Performance metrics (accuracy, precision, recall, F1 score, and ROC AUC) of seven machine learning models predicting 2-year DFS

	ROC AUC	Accuracy	Precision	Recall	F1 score
Decision tree	0.533	0.507	0.781	0.328	0.462
Random forest	0.664	0.630	0.750	0.638	0.689
Bagging	0.629	0.607	0.693	0.701	0.697
AdaBoost	0.616	0.630	0.740	0.655	0.695
GradientBoost	0.668	0.681	0.797	0.678	0.733
XGBoost	0.652	0.630	0.652	0.914	0.761
DNN	0.605	0.641	0.718	0.730	0.724

Evaluated on test set (10%). DFS, disease-free survival; DNN, deep neural network; ROC AUC, receiver operating characteristic area under curve.

Table 3.

Univariable Cox regression analysis showing HR, 95% CI, and p-values for key features affecting DFS

Feature	HR (95% CI)	p-value
Age	1.014 (1.008-1.019)	< 0.001
Sex (reference: female)
Male	1.419 (1.227-1.641)	< 0.001
ECOG PS	1.104 (0.855-1.426)	0.447
Baseline Hb	0.976 (0.921-1.034)	0.412
Baseline PLT	1.001 (1.000-1.002)	0.139
Baseline LDH	1.003 (1.002-1.005)	< 0.001
Baseline CRP	1.005 (1.002-1.009)	0.005
Baseline NLR	1.173 (1.110-1.240)	< 0.001
Histological type (reference: acral lentiginous)
Superficial spreading	0.846 (0.635-1.129)	0.256
Nodular	1.700 (1.307-2.211)	< 0.001
Lentigo maligna	0.365 (0.091-1.464)	0.155
Ocular	0.442 (0.367-0.533)	< 0.001
Mucosal	2.564 (2.139-3.074)	< 0.001
Baseline BMI (reference: normal)
Underweight	0.993 (0.643-1.532)	0.973
Overweight	0.694 (0.575-0.837)	< 0.001
Obese	2.323 (1.984-2.719)	< 0.001

BMI, body mass index; CI, confidence interval; CRP, C-reactive protein; DFS, disease-free survival; ECOG PS, Eastern Cooperative Oncology Group performance status; Hb, hemoglobin; HR, hazard ratio; LDH, lactate dehydrogenase; NLR, neutrophil-to-lymphocyte ratio; PLT, platelet count.

Table 4.

Median DFS, 2-year DFS rate, and log-rank test results

Feature	Median DFS (mo) (95% CI)	2-Year DFS rate (%) (95% CI)	Log-rank p-value
Total	63.93 (52.93-85.33)	67.65 (65.16-70.00)
Sex			< 0.005
Female	98.97 (71.23-131.27)	72.50 (69.03-75.39)
Male	42.13 (36.53-54.53)	62.19 (58.40-65.73)
Primary site			< 0.005
Cutaneous	57.67 (48.03-68.33)	66.80 (63.54-69.84)
Ocular	165.23 (134.93-inf)	82.94 (78.76-86.37)
Mucosal	18.37 (13.90-22.40)	40.77 (33.75-47.66)
Histological type			< 0.005
Acral lentiginous	Not reached	77.75 (72.54-82.09)
Superficial spreading	103.93 (45.47-inf)	72.45 (61.42-77.73)
Nodular	30.70 (17.57-45.40)	53.34 (41.61-61.71)
Lentigo maligna	Not reached	90.00 (47.30-98.53)
Ocular	165.23 (134.93-inf)	82.94 (78.76-86.37)
Mucosal	18.37 (13.90-22.40)	40.77 (33.75-47.66)
Pathological stage			< 0.005
In situ	Not reached	96.82 (91.70-98.80)
I	132.80 (126.50-inf)	89.35 (85.01-92.49)
II	43.07 (33.13-52.93)	62.76 (57.78-67.33)
III	19.37 (17.23-24.83)	44.60 (37.86-51.11)
Baseline BMI			< 0.005
Underweight	59.60 (30.87-inf)	71.98 (55.88-83.05)
Normal	97.57 (68.90-127.80)	71.06 (67.75-74.09)
Overweight	108.70 (85.33-150.60)	75.81 (70.68-80.17)
Obese	22.83 (19.00-27.87)	48.58 (42.72-54.19)

BMI, body mass index; CI, confidence interval; DFS, disease-free survival.

Table 5.

Comparison of our proposed TNM-free ML model with prior melanoma survival prediction studies

Study	Population	TNM used	Model type	Outcome	Metrics
This study	Korean, stage I-III melanoma	X	XGBoost	2-yr DFS	F1=0.761
Cozzolino et al. (2023) [21]	US, all stages	O	Deep neural network	5-yr OS	F1=0.800
Wan et al. (2022) [22]	Chinese, all stages	O	GradientBoost	3-yr OS	AUC=0.76

Key aspects include population, TNM variable usage, model type, prediction outcome, and reported performance metrics (F1 score or area under the receiver operating characteristic curve [AUC]). DFS, disease-free survival; ML, machine learning; OS, overall survival.

REFERENCES

1. Siegel RL, Kratzer TB, Giaquinto AN, Sung H, Jemal A. Cancer statistics, 2025. CA Cancer J Clin. 2025;75:10–45. Article PubMed PMC
2. Korean Central Cancer Registry, Ministry of Health and Welfare. Annual report of cancer statistics in Korea in 2024. Ministry of Health and Welfare; 2025
3. Gershenwald JE, Scolyer RA, Hess KR, Sondak VK, Long GV, Ross MI, et al. Melanoma staging: evidence-based changes in the American Joint Committee on Cancer eighth edition cancer staging manual. CA Cancer J Clin. 2017;67:472–92. Article PubMed PMC PDF
4. Olsen CM, Whiteman DC. Risk stratification for melanoma. Oncotarget. 2019;10:1868–9. Article PubMed PMC
5. Balch CM, Gershenwald JE, Soong SJ, Thompson JF, Atkins MB, Byrd DR, et al. Final version of 2009 AJCC melanoma staging and classification. J Clin Oncol. 2009;27:6199–206. Article PubMed PMC
6. Kim SY, Yun SJ. Cutaneous melanoma in Asians. Chonnam Med J. 2016;52:185–93. Article PubMed PMC PDF
7. Bradford PT, Goldstein AM, McMaster ML, Tucker MA. Acral lentiginous melanoma: incidence and survival patterns in the United States, 1986-2005. Arch Dermatol. 2009;145:427–34. PubMed PMC
8. Jung HJ, Kweon SS, Lee JB, Lee SC, Yun SJ. A clinicopathologic analysis of 177 acral melanomas in Koreans: relevance of spreading pattern and physical stress. JAMA Dermatol. 2013;149:1281–8. Article PubMed
9. Ravaioli GM, Dika E, Lambertini M, Chessa MA, Fanti PA, Patrizi A. Acral melanoma: correlating the clinical presentation to the mutational status. G Ital Dermatol Venereol. 2019;154:567–72. Article PubMed
10. Asgari MM, Shen L, Sokil MM, Yeh I, Jorgenson E. Prognostic factors and survival in acral lentiginous melanoma. Br J Dermatol. 2017;177:428–35. Article PubMed PDF
11. Kavitha P, Ayyappan G, Jayagopal P, Mathivanan SK, Mallik S, Al-Rasheed A, et al. Detection for melanoma skin cancer through ACCF, BPPF, and CLF techniques with machine learning approach. BMC Bioinformatics. 2023;24:458.Article PubMed PMC PDF
12. Li L, Zhang Q, Ding Y, Jiang H, Thiers BH, Wang JZ. Automatic diagnosis of melanoma using machine learning methods on a spectroscopic system. BMC Med Imaging. 2014;14:36.Article PubMed PMC PDF
13. Figueroa-Silva O, Pastur Romay LA, Viruez Roca RD, Rojas MD, Suarez-Penaranda JM. Machine learning techniques in predicting BRAF mutation status in cutaneous melanoma from clinical and histopathologic features. Appl Immunohistochem Mol Morphol. 2022;30:674–80. Article PubMed
14. Bibault JE, Giraud P, Burgun A. Big Data and machine learning in radiation oncology: state of the art and future prospects. Cancer Lett. 2016;382:110–7. Article PubMed
15. Amaral T, Chatziioannou E, Nuebling A, Nanz L, Sinnberg T, Niessner H, et al. Risk stratification using the Merlin Assay (CP-GEP) in an independent cohort of 930 patients with clinical stage I/II melanoma who did not undergo sentinel lymph node biopsy. Eur J Cancer. 2025;220:115372.Article PubMed
16. Grant S, Andrew T, Mullin S, Podlipnik S, Puig S, McArthur G, et al. The development of a risk stratification tool for localized melanomas through the integrated application of supervised and unsupervised machine learning techniques. J Am Acad Dermatol. 2025;93:1321–4. Article PubMed
17. Chen T, Guestrin C. XGBoost: a scalable tree boosting system. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining; 2016 Aug 13-17; San Francisco, CA, USA. New York: Association for Computing Machinery; 2016. p. 785-94
18. DeVries Z, Locke E, Hoda M, Moravek D, Phan K, Stratton A, et al. Using a national surgical database to predict complications following posterior lumbar surgery and comparing the area under the curve and F1-score for the assessment of prognostic capability. Spine J. 2021;21:1135–42. Article
19. Esteva A, Kuprel B, Novoa RA, Ko J, Swetter SM, Blau HM, et al. Dermatologist-level classification of skin cancer with deep neural networks. Nature. 2017;542:115–8. PubMed PMC
20. Kawahara J, BenTaieb A, Hamarneh G. Deep features to classify skin lesions. In: Proceedings of the 2016 IEEE 13th International Symposium on Biomedical Imaging (ISBI); 2016 Apr 13-16; Prague, Czech Republic. Piscataway: Institute of Electrical and Electronics Engineers; 2016
21. Cozzolino C, Buja A, Rugge M, Miatton A, Zorzi M, Vecchiato A, et al. Machine learning to predict overall short-term mortality in cutaneous melanoma. Discov Oncol. 2023;14:13.Article PubMed PMC PDF
22. Wan G, Nguyen N, Liu F, DeSimone MS, Leung BW, Rajeh A, et al. Prediction of early-stage melanoma recurrence using clinical and histopathologic features. NPJ Precis Oncol. 2022;6:79.Article PubMed PMC PDF
23. Hahn HM, Lee KG, Choi W, Cheong SH, Myung KB, Hahn HJ. An updated review of mucosal melanoma: survival meta-analysis. Mol Clin Oncol. 2019;11:116–26. Article PubMed PMC
24. Bishop KD, Olszewski AJ. Epidemiology and survival outcomes of ocular and mucosal melanomas: a population-based analysis. Int J Cancer. 2014;134:2961–71. Article PubMed
25. Tan LL, Hong J, Goh WL, Chang EW, Yang VS, Poon E, et al. Clinical features and survival outcomes of ocular melanoma in a multi-ethnic Asian cohort. Sci Rep. 2020;10:16367.Article PubMed PMC PDF
26. Jung JM, Jung CJ, Won CH, Chang SE, Lee MW, Choi JH, et al. Different progression pattern between acral and nonacral melanoma: a retrospective, comparative, clinicoprognostic study of 492 cases of primary cutaneous melanoma according to tumor site. Indian J Dermatol Venereol Leprol. 2021;87:498–508. Article PubMed
27. Mandala M, Rutkowski P, Galli F, Patuzzo R, De Giorgi V, Rulli E, et al. Acral lentiginous melanoma histotype predicts outcome in clinical stage I-II melanoma patients: an international multicenter study. ESMO Open. 2022;7:100469.Article PubMed PMC
28. Park SM, Jeong HY, Kim HS, Kwak MJ, Yi CC, Bae YC. Epidemiology and prognostic factors of malignant melanoma among 200 Asian patients from a single medical center. Indian J Surg. 2024;86:310–7. Article PDF
29. Thompson JF, Soong SJ, Balch CM, Gershenwald JE, Ding S, Coit DG, et al. Prognostic significance of mitotic rate in localized primary cutaneous melanoma: an analysis of patients in the multi-institutional American Joint Committee on Cancer melanoma staging database. J Clin Oncol. 2011;29:2199–205. PubMed PMC
30. Deckers EA, Kruijff S, Bastiaannet E, van Ginkel RJ, Hoekstra-Weebers JE, Hoekstra HJ. Obesity is not associated with disease-free interval, melanoma-specific survival, or overall survival in patients with clinical stage IB-II melanoma after SLNB. J Surg Oncol. 2021;124:655–64. Article PubMed PMC PDF
31. Jansen J, Garmyn M, Guvenc C. The effect of body mass index on melanoma biology, immunotherapy efficacy, and clinical outcomes: a narrative review. Int J Mol Sci. 2024;25:6433.Article PubMed PMC
32. Shwartz-Ziv R, Armon A. Tabular data: deep learning is not all you need. Inform Fusion. 2022;81:84–90. Article

Figure & Data

REFERENCES

Citations

Citations to this article as recorded by

PubReader
ePub Link

Cite

CITE: export Copy Download Format; Close

Download Citation

Download a citation file in RIS format that can be imported by all major citation management software, including EndNote, ProCite, RefWorks, and Reference Manager.

Format:

RIS — For EndNote, ProCite, RefWorks, and most other reference management software
BibTeX — For JabRef, BibDesk, and other BibTeX-specific software

Include:

Citation for the content below
Citation and abstract for the content below

Risk Prediction Based on Clinicopathologic Features in Korean Melanoma Patients by Machine Learning

DOI: https://doi.org/10.4143/crt.2025.1154

XML Download

Abstract
Introduction
Materials and Methods
Results
Discussion
Electronic Supplementary Material
NOTES
REFERENCES

Risk Prediction Based on Clinicopathologic Features in Korean Melanoma Patients by Machine Learning

Fig. 1. Feature importances of decision tree (A), random forest (B), AdaBoost (C), GradientBoost (D), and XGBoost (E), computed from model-specific metrics such as Gini impurity or information gain. BMI, body mass index; CRP, C-reactive protein; ECOG, Eastern Cooperative Oncology Group; Hb, hemoglobin; LDH, lactate dehydrogenase; NLR, neutrophil-to-lymphocyte ratio; PLT, platelet count.

Fig. 2. Kaplan-Meier curves of patients compared by sex (A), pathological stage (B), primary site (C), histological type (D), and baseline body mass index (BMI) (E). All survival curves were estimated using the ‘lifelines’ package, and significance was evaluated via log-rank test.

Fig. 1.

Fig. 2.

Risk Prediction Based on Clinicopathologic Features in Korean Melanoma Patients by Machine Learning

	Total	Cutaneous					Ocular	Mucosal
	Total	Total	Acral lentiginous	Superficial spreading	Nodular	Lentigo maligna	Ocular	Mucosal
Total	1,657	1,015 (61.3)	365 (22.0)	160 (9.7)	104 (6.3)	15 (0.9)	429 (25.9)	213 (12.9)
Age (yr)
Median (range)	59 (19-97)	60 (19-95)	63 (21-93)	58 (19-87)	58 (24-87)	70 (42-86)	56 (19-90)	62 (28-97)
Sex
Male	778 (47.0)	469 (46.2)	171 (46.8)	70 (43.8)	52 (50.0)	10 (66.7)	206 (48.0)	103 (48.4)
Female	879 (53.0)	546 (53.8)	194 (53.2)	90 (56.3)	52 (50.0)	5 (33.3)	223 (52.0)	110 (51.6)
ECOG PS
0	841 (50.8)	575 (56.7)	254 (69.6)	121 (75.6)	68 (65.4)	12 (80.0)	173 (40.3)	93 (43.7)
1	90 (5.4)	18 (1.8)	3 (0.8)	2 (1.3)	1 (1.0)	0	54 (12.6)	18 (8.5)
2	7 (0.4)	4 (0.4)	1 (0.3)	0	0	0	1 (0.2)	2 (0.9)
3	2 (0.1)	1 (0.1)	0	0	0	0	0	1 (0.5)
Pathological stage
In situ	177 (10.7)	174 (17.1)	96 (26.3)	17 (10.6)	0	6 (40.0)	0	3 (1.4)
I	321 (19.4)	242 (23.8)	102 (28.0)	66 (41.3)	13 (12.5)	6 (40.0)	66 (15.4)	13 (6.1)
II	448 (27.0)	274 (27.0)	111 (30.4)	43 (26.9)	47 (45.2)	1 (6.7)	153 (35.7)	21 (9.9)
III	236 (14.2)	177 (17.4)	48 (13.2)	29 (18.1)	40 (38.5)	0	30 (7.0)	29 (13.6)
Baseline BMI
Underweight	45 (2.7)	25 (2.5)	6 (1.6)	5 (3.1)	2 (1.9)	0	15 (3.5)	5 (2.3)
Normal	939 (56.7)	554 (54.6)	217 (59.5)	109 (68.1)	50 (48.1)	12 (80.0)	279 (65.0)	106 (49.8)
Overweight	362 (21.9)	209 (20.6)	92 (25.2)	30 (18.8)	26 (25.0)	2 (13.3)	104 (24.2)	49 (23.0)
Obese	311 (18.8)	227 (22.4)	50 (13.7)	16 (10.0)	26 (25.0)	1 (6.7)	31 (7.2)	53 (24.9)
Relapse
No	1,102 (66.5)	658 (64.8)	270 (74.0)	116 (72.5)	46 (44.2)	13 (86.7)	334 (77.9)	110 (51.6)
Yes	555 (33.5)	357 (35.2)	95 (26.0)	44 (27.5)	58 (55.8)	2 (13.3)	95 (22.1)	103 (48.4)
Death
No	1,098 (66.3)	689 (67.9)	293 (80.3)	130 (81.3)	66 (63.5)	14 (93.3)	315 (73.4)	94 (44.1)
Yes	559 (33.7)	326 (32.1)	72 (19.3)	30 (18.8)	38 (36.5)	1 (6.7)	114 (26.6)	119 (55.9)

	ROC AUC	Accuracy	Precision	Recall	F1 score
Decision tree	0.533	0.507	0.781	0.328	0.462
Random forest	0.664	0.630	0.750	0.638	0.689
Bagging	0.629	0.607	0.693	0.701	0.697
AdaBoost	0.616	0.630	0.740	0.655	0.695
GradientBoost	0.668	0.681	0.797	0.678	0.733
XGBoost	0.652	0.630	0.652	0.914	0.761
DNN	0.605	0.641	0.718	0.730	0.724

Feature	HR (95% CI)	p-value
Age	1.014 (1.008-1.019)	< 0.001
Sex (reference: female)
Male	1.419 (1.227-1.641)	< 0.001
ECOG PS	1.104 (0.855-1.426)	0.447
Baseline Hb	0.976 (0.921-1.034)	0.412
Baseline PLT	1.001 (1.000-1.002)	0.139
Baseline LDH	1.003 (1.002-1.005)	< 0.001
Baseline CRP	1.005 (1.002-1.009)	0.005
Baseline NLR	1.173 (1.110-1.240)	< 0.001
Histological type (reference: acral lentiginous)
Superficial spreading	0.846 (0.635-1.129)	0.256
Nodular	1.700 (1.307-2.211)	< 0.001
Lentigo maligna	0.365 (0.091-1.464)	0.155
Ocular	0.442 (0.367-0.533)	< 0.001
Mucosal	2.564 (2.139-3.074)	< 0.001
Baseline BMI (reference: normal)
Underweight	0.993 (0.643-1.532)	0.973
Overweight	0.694 (0.575-0.837)	< 0.001
Obese	2.323 (1.984-2.719)	< 0.001

Feature	Median DFS (mo) (95% CI)	2-Year DFS rate (%) (95% CI)	Log-rank p-value
Total	63.93 (52.93-85.33)	67.65 (65.16-70.00)
Sex			< 0.005
Female	98.97 (71.23-131.27)	72.50 (69.03-75.39)
Male	42.13 (36.53-54.53)	62.19 (58.40-65.73)
Primary site			< 0.005
Cutaneous	57.67 (48.03-68.33)	66.80 (63.54-69.84)
Ocular	165.23 (134.93-inf)	82.94 (78.76-86.37)
Mucosal	18.37 (13.90-22.40)	40.77 (33.75-47.66)
Histological type			< 0.005
Acral lentiginous	Not reached	77.75 (72.54-82.09)
Superficial spreading	103.93 (45.47-inf)	72.45 (61.42-77.73)
Nodular	30.70 (17.57-45.40)	53.34 (41.61-61.71)
Lentigo maligna	Not reached	90.00 (47.30-98.53)
Ocular	165.23 (134.93-inf)	82.94 (78.76-86.37)
Mucosal	18.37 (13.90-22.40)	40.77 (33.75-47.66)
Pathological stage			< 0.005
In situ	Not reached	96.82 (91.70-98.80)
I	132.80 (126.50-inf)	89.35 (85.01-92.49)
II	43.07 (33.13-52.93)	62.76 (57.78-67.33)
III	19.37 (17.23-24.83)	44.60 (37.86-51.11)
Baseline BMI			< 0.005
Underweight	59.60 (30.87-inf)	71.98 (55.88-83.05)
Normal	97.57 (68.90-127.80)	71.06 (67.75-74.09)
Overweight	108.70 (85.33-150.60)	75.81 (70.68-80.17)
Obese	22.83 (19.00-27.87)	48.58 (42.72-54.19)

Study	Population	TNM used	Model type	Outcome	Metrics
This study	Korean, stage I-III melanoma	X	XGBoost	2-yr DFS	F1=0.761
Cozzolino et al. (2023) [21]	US, all stages	O	Deep neural network	5-yr OS	F1=0.800
Wan et al. (2022) [22]	Chinese, all stages	O	GradientBoost	3-yr OS	AUC=0.76

Table 1. Baseline demographic and clinicopathologic characteristics of the study cohort

Values are presented as number (%) unless otherwise indicated. BMI, body mass index; ECOG PS, Eastern Cooperative Oncology Group performance status.

Table 2. Performance metrics (accuracy, precision, recall, F1 score, and ROC AUC) of seven machine learning models predicting 2-year DFS

Evaluated on test set (10%). DFS, disease-free survival; DNN, deep neural network; ROC AUC, receiver operating characteristic area under curve.

Table 3. Univariable Cox regression analysis showing HR, 95% CI, and p-values for key features affecting DFS

Table 4. Median DFS, 2-year DFS rate, and log-rank test results

BMI, body mass index; CI, confidence interval; DFS, disease-free survival.

Table 5. Comparison of our proposed TNM-free ML model with prior melanoma survival prediction studies