Thibault Anani

In this paper, we study the problem of feature selection with the aim of improving the predictive performance of machine-learning models. Wang et al. (2013) showed that Differential Evolution (DE), a metaheuristic comparatively underused in both the academic literature and industry, can outperform more widespread approaches such as Particle Swarm Optimisation and Genetic Algorithms. However, their results have several limitations : no comparison with filter-based methods or greedy algorithms, and evaluation solely on artificial score functions rather than datasets.We present a broader study that includes filter methods, greedy heuristics and other metaheuristics, together with a suite of datasets spanning both synthetic benchmarks and real-world cases. Our results confirm that DE is frequently competitive and often superior in predictive performance and robustness. More significantly, an ablation over DE operators hints that a specific mutation strategy, not the one adopted in Wang et al., consistently yields stronger and more stable results across learners and datasets. Building on these findings, we introduce TiDE, a DE-based variant that improves search efficiency and stability across learners. We further propose TiDE-OPT, which tunes a key hyperparameter to adapt to dataset characteristics. Across datasets and learners, TiDE and TiDE-OPT achieve top-tier performance, with TiDE-OPT most often attaining the best results, while maintaining compact feature subsets.

Background: This study focuses on Amyotrophic Lateral Sclerosis (ALS), a progressive neurodegenerative disease that affects motor neurons. ALS patients suffer from motor weakness, atrophy, spasticity, and difficulties in speaking, swallowing and breathing. Predicting the survival and progression of ALS is essential for optimising patient care, interventions and informed decision-making. Methods: Using the PRO-ACT (Pooled Resource Open-Access ALS Clinical Trials) database and clinical trial data from Exonhit Therapeutics, we applied machine learning techniques to predict disease progression based on ALS Functional Rating Scale (ALSFRS) scores and patient survival over one year of follow-up. Our models were validated through 10-fold cross-validation. Kaplan-Meier estimation was used to cluster patients effectively according to their profiles. To enhance the predictive accuracy of our models, we performed feature selection prior to analysis. Results: Logistic regression combined with feature selection yielded a Balanced Accuracy score of 76% (68.6% to 79.8% in each fold) on validation data and 76.33% on test data. Integrating model results with a Kaplan-Meier estimates, we identified four patient clusters, with a C-Index of 0.81 and an overall log-rank test p-value ≤ 0.0001. Our method demonstrated high accuracy in predicting ALS-FRS score progression at 3 months (RMSE = 2.88 and adjusted R² = 0.784 with random forest). This study showcases the potential of machine learning models to provide significant predictive insights in ALS, enhancing the understanding of disease dynamics and supporting patient care.

This study addresses Amyotrophic Lateral Sclerosis (ALS), a progressive neurodegenerative disease affecting motor neurons. ALS patients experience motor weakness, atrophy, spasticity, and difficulty in speech, swallowing, and breathing. Predicting ALS survival and progression is crucial for tailored care, interventions, and informed decision-making. Utilising the PRO-ACT database and Exonhit Therapeutics' clinical trial data, machine learning may allow predicting disease progression based on ALS Functional Rating Scale (ALSFRS) scores and patient survival within one year of follow-up. To comprehensively understand patient profiles, unsupervised learning via UMAP was applied for dimension reduction, effectively grouping patients based on features. To optimise the predictive quality of our models, we selected features prior to the analysis process. Ridge regression combined with feature selection yielded a predictive quality of 80.32\% for patients' survival and UMAP achieved prediction rates from 67.93\% when considering all clusters to 92.75\% when focusing on the most homogeneous ones, emphasising the relevance of feature selection. Our method showed accurate performance in predicting ALSFRS score progression at 3 months (RMSE = 2.86 and Adjusted R² = 0.796). This study demonstrates the ability of the models to provide meaningful predictive insights in the specific context of ALS, enhancing the understanding of disease dynamics and facilitating personalised patient care.