Lightweight Multimodal Deep Learning Framework for Cerebellar Ataxia Detection Using an MLP Encoder and Hybrid Optimizer

Background Cerebellar ataxia is characterized by abnormalities in gait, speech, ocular movement, and coordination, making accurate detection challenging. Purpose This study develops a lightweight multimodal Deep Learning (DL)-based framework for cerebellar ataxia detection. Research Design The proposed approach integrates voice, gaze, and body pose features from independent public datasets and performs class-level multimodal feature fusion to capture complementary motor and behavioral cues. Fine-grained features such as Mel Frequency Cepstral Coefficients (MFCCs) and vocal perturbation measures, gaze fixation dynamics, and spatiotemporal body pose features are extracted using Python-compatible tools. Each modality is processed with a specific Multilayer Perceptron (MLP)-based encoder to produce small embeddings, which are fused through feature concatenation and processed by a fully connected MLP classifier. To ensure stable convergence and robust generalization capability on limited and noisy clinical data, a novel hybrid optimizer, Adaptive Sharp-Lookahead Lite, is used. Results Experimental results demonstrate that the proposed framework achieves an accuracy of 99.56% in distinguishing ataxia from normal cases under the adopted experimental setting. Conclusions The findings demonstrate the potential of the framework as a lightweight AI-based screening approach for resource-constrained settings.