Exploring risk factors for amyotrophic lateral sclerosis through machine learning approaches

Objective: This study aimed to identify determining factors in amyotrophic lateral sclerosis (ALS) using various machine learning (ML) models and to systematically compare their classification performance. Methods: Data from ALS patients and controls provided by the OnWebDUALS project were analyzed within the CRISP-DM framework. Four datasets addressing missing values were examined with ML models, and the dataset yielding the best performance was further subjected to feature selection for model evaluation. Classification results were assessed using Accuracy, ROC-AUC, PR-AUC, F1-score, Precision, Recall, and MCC metrics, alongside feature importance plots. A clinical decision support system (CDSS) was developed based on the best-performing model and features, with interpretability enhanced through SHAP and LIME methods. Results: The analysis included 53 categorical variables from 1,142 ALS patients and 821 controls. ML models achieved their highest performance on the dataset where observations with missing values were excluded. Feature selection with XGB, LGBM, and CB reduced the 53 initial variables to 37 largely overlapping features, with intersection and union sets also explored. Among the ML models, boosting algorithms outperformed others, with LGBM showing the best results (Accuracy: 0.82, ROC-AUC: 0.88, PR-AUC: 0.85, F1-score: 0.82, Precision: 0.80, Recall: 0.84, MCC: 0.64). The most critical predictors were identified as gender, occupation, smoking, country, and hypertension. Conclusion: In the categorical ALS dataset, excluding incomplete cases yielded the highest performance. While feature selection algorithms identified the same number of variables, their prioritizations varied. The best results were achieved with LGBM trained on 37 features selected by XGB. Although not intended for direct diagnosis, the developed CDSS provides data-driven outputs for ALS classification, with potential to assist clinicians and contribute to future research.