Visualising and quantifying relevant parkinsonian gait patterns using 3D convolutional network

•A markerless strategy to support diagnosis of Parkinson’s disease from gait videos.•Spatiotemporal saliency maps that stand out parkinsonian patterns in gait videos.•The model highlights spatiotemporal patterns of legs during the single support phase.•The approach achieves 94.89% of accuracy in Parkinson’s disease classification.