Machine learning and logistic regression in estimating survival in patients with high-malignant deep-seated soft tissue sarcomas: development and analysis based on a population-based retrospective cohort

TL;DR

This study compares machine learning and logistic regression models to predict 5-year survival in patients with high-grade soft tissue sarcomas, finding that logistic regression performs better in this population-based cohort.

Contribution

The study develops and evaluates ML models for survival prediction in soft tissue sarcomas using a modern Scandinavian population-based dataset.

Findings

Logistic regression outperformed random forest in AUC, sensitivity, and specificity for 5-year survival prediction.

Trunk location, grade 3 tumors, and early chemotherapy were identified as key negative predictors of survival.

Random forest showed better performance during training but underperformed after internal validation.

Abstract

Soft tissue sarcomas are a heterogeneous group of malignant tumors with a high risk of metastasis, primarily to the lungs, making accurate survival prediction an essential part of long-term planning. No machine learning (ML) survival prediction models have been developed using a modern, population-based dataset from Scandinavia. We aimed to develop and compare ML models with logistic regression in predicting 5-year survival in soft tissue sarcoma patients and identify key predictive variables.

This retrospective cohort study included patients diagnosed with deep-seated, high-grade soft tissue sarcomas of the extremities and trunk wall in Denmark from 2000 to 2016. Logistic regression was compared with 4 developed ML models, including random forest. Performance was assessed using the area under the curve (AUC), sensitivity, specificity, and calibration metrics, with a 70:30 training–test split and 5-fold cross-validation to evaluate the models.

516 patients were included, of whom 226 (44%) died within 5 years following surgery. Random forest demonstrated the best ML performance on the training set and was compared with logistic regression on the test set. Logistic regression achieved an AUC of 0.74 (95% confidence interval [CI] 0.66–0.82), outperforming random forest‘s AUC of 0.65 (CI 0.56–0.74). Logistic regression also had higher sensitivity (0.65 vs 0.59) and specificity (0.72 vs 0.69), while random forest had a lower Brier score (0.38 vs 0.41).

Although the developed random forest ML model performed well during training, logistic regression outperformed it after internal validation. Soft tissue sarcomas located in the trunk, grade 3 tumors, and chemotherapy within 3 months of surgery demonstrated the highest negative effect on survival, consistent with current treatment protocols in which patients with high-risk disease are managed with more aggressive multimodal therapy. Further external validation and assessment of clinical utility are required before potential clinical implementation.

Figures