Hualou Liang and Felix Agbavor Win 1st and 2nd Place at the INTERSPEECH 2024 TAUKADIAL Challenge for Dementia Prediction Using LLMs
November 11, 2024
Hualou Liang, PhD, professor, and Felix Agbavor, PhD candidate (Advisor: H. Liang), both in the School of Biomedical Engineering, Science and Health Systems, participated in the INTERSPEECH 2024 TAUKADIAL Challenge, an international competition for dementia prediction using Large Language Models (LLMs) for speech-based cognitive assessment in Chinese and English. The challenge has two tasks: (1) MCI Detection: a classification task, where participants create models to distinguish healthy control speech from Mild Cognitive Impairment (MCI) speech, and (2) MMSE Prediction: a cognitive test score prediction (regression) task, where a model is created to infer the subject's Mini Mental Status Examination (MMSE) or Montreal Cognitive Assessment (MoCA) scores based on connected (spontaneous) speech data. In both tasks, Dr. Liang and Felix won overwhelmingly. Drexel ranked 2nd for MCI Detection (Task 1) and 1st for MMSE Prediction (Task 2).
They have since been granted a US patent and awarded an NIH grant via PennAITech. For this competition, they developed innovative on-prem open-source Small Language Models (SLMs), rather than LLMs, to address the privacy and security concerns raised by their clinical collaborators at Jefferson University. Moreover, they are exploring models that generalize across languages (English and Chinese), and not just focusing on English speech, as most current studies have done. Their new algorithm fared rather well in the international community and far surpassed those at some top US universities. Their work has caught the attention of FDA, with their issuing an exclusive call for proposal for Drexel. The project has now been funded for the Biopharmaceutics Classification System (BCS) drug classification using LLMs, as the transformative potential of artificial intelligence (AI), particularly LLMs, for digital health continues to advance.