A Novel Predictive Algorithm Design to Identify Amyotrophic Lateral Sclerosis (ALS) by using Elevated Hybrid Deep Learning Strategy

The motor functions alongside speech experience major deterioration due to the progressive nature of Amyotrophic Lateral Sclerosis (ALS). The proper identification of the disease during early stages enables both disease control and better patient results. The research proposes a Novel Predictive Algorithm Design to Identify ALS by applying Elevated Hybrid Deep Learning Strategy on a Kaggle dataset containing 131 voice features derived from sustained vowel phonations. The proposed algorithm incorporates CNN to extract deep features then uses LightGBM for structured analysis which finally combines both results at the feature level to boost classification performance. The developed model exhibited high accuracy at 94.5% along with precision of 92.3% and recall value of 95.8% while achieving an AUC-ROC score of 96.7%. These results outperform traditional models like Support Vector Machines (87.2%), Random Forest (89.8%), and even standalone CNNs (92.7%). SHAP-based feature importance analysis identified three crucial vocalization features including Jitter, Shimmer and Harmonic-to-Noise Ratios in detecting the presence of ALS. The model demonstrated excellent robustness through routing its data through Stratified K-Fold Cross-Validation which produced consistent performance in each iteration. This combination structure shows promise for non-contact voice diagnosis of ALS which advances both medical acoustics and AI-based healthcare technologies.