Currently, the detection and assessment of Parkinson's Disease (PD) continue to be a challenge for modern medicine. This study investigated the possibility of implementing a multimodal audio-visual deep-learning model to detect PD. To begin, we designed a bidirectional long short-term memory model (BiLSTM) under different considerations. After comparing the model combinations with their unimodal counterparts, we conducted a three-way repeated measures analysis of variance (RMANOVA) and multiple pairwise t-tests to validate the following results. First, the audio-visual models performed significantly better than the unimodal models. However, an outlier in the performance showed the instability of the models when they were trained on insufficient data. Second, the models only compensate for the disturbances if both modalities are good quality. Finally, the models work better on raw and low-level features than handcrafted ones.