Developing Predictive Models for Rare Diseases Clinical Trials

Abstract

Research on rare disorders, which are mostly inherited and often exhibit neurological symptoms, is gravely under-represented. This research uses gait analysis to categorize patients with hereditary cerebellar ataxia versus healthy controls, and it builds a predictive modelling framework for therapeutic trials targeting uncommon disorders. This study presents a gait classification methodology for distinguishing individuals with hereditary cerebellar ataxia from healthy subjects, leveraging an inertial measurement unit (IMU) to capture gait metrics and trunk acceleration from thirty diagnosed individuals and one hundred healthy participants. The dataset undergoes preprocessing steps, including outlier correction using interquartile range (IQR). The Random Forest (RF) classifier ranks features by importance, revealing significant predictors like CV-step-length, sLLEAP, and HRAP. To address class imbalance, various data augmentation techniques—under-sampling, over-sampling, SMOTE, GAN, and CT-GAN—are employed. Among these, CT-GAN exhibits the highest effectiveness in producing balanced data, contributing to enhanced classification accuracy. Among a performance parameter utilised to assess an effectiveness of a model, CT-GAN achieved a 90% accuracy rate, an 88% recall rate, an 88% F1-score, and a 0.90 ROC-AUC. The results underscore the advantages of CT-GAN in improving model performance for imbalanced clinical data, especially in rare disease prediction.