Interpretable models for clinical risk

Research approach

Six classification methods were compared using a pipeline that addressed class imbalance. SHAP analysis explored which features influenced predictions.

Key findings

The paper reports an SVM ROC-AUC of approximately 0.79 and a peak KNN recall of 0.80; these describe different models and metrics.
Glucose and C-reactive protein were important predictors in the reported analysis.
Laboratory-marker models produced a higher amputation-class F1-score than wound-only models in the feature-set comparison below.

Predicting the amputation class

F1-score by feature set · optimized SVM comparison

Wound characteristics

0.37

Metabolic & inflammatory markers

0.49

0F1-score (0–1)1

Original portfolio chart using the amputation-class F1-scores reported in Table 3 of Menon et al., Physica Scripta (2026), DOI: 10.1088/1402-4896/ae4ae5. Test support: 73 observations. These are feature-set comparisons, not overall model-performance estimates.

Publication

Predicting lower limb amputation risk based on clinical factors and wound characteristics using machine learning

Gayathry R Menon, Sayooj Aby Jose, Sweetymol Jose, Jobin K Thomas & Anuwat Jirawattanapanit

Physica Scripta · 101, 105004

Read the full paper

Summary prepared from the linked publication. The full article provides the complete methods, results, and limitations.