Abstract
BACKGROUND/OBJECTIVES: Parkinson's disease (PD) affects more than 6 million people worldwide. Its accurate diagnosis and monitoring are key factors to reduce its economic burden. Typical approaches consider either speech signals or video recordings of the face to automatically model abnormal patterns in PD patients. METHODS: This paper introduces, for the first time, a new methodology that performs the synchronous fusion of information extracted from speech recordings and their corresponding videos of lip movement, namely the bimodal approach. RESULTS: Our results indicate that the introduced method is more accurate and suitable than unimodal approaches or classical asynchronous approaches that combine both sources of information but do not incorporate the underlying temporal information. CONCLUSIONS: This study demonstrates that using a synchronous fusion strategy with concatenated projections based on attention mechanisms, i.e., speech-to-lips and lips-to-speech, exceeds previous results reported in the literature. Complementary information between lip movement and speech production is confirmed when advanced fusion strategies are employed. Finally, multimodal approaches, combining visual and speech signals, showed great potential to improve PD classification, generating more confident and robust models for clinical diagnostic support.