Thread by @iam_mahajan

I'm very much excited to share that our with Prof. @baths_veeky on Acoustic and Language-based deep learning approaches to Alzheimer's dementia detection from spontaneous speech has now been published at Frontiers in #Aging #Neuroscience @FrontNeurosci!
frontiersin.org/articles/10.33…

Similar to how RNNs are trained on word embeddings and POS-tags for language based tasks, we propose recurrent stream which processes paralinguistic acoustic features and looks for temporal patterns relevant to early detection of dementia.

We show that enriching a recurrent feature processing stream with dementia specific features such as idea-density, disfluency, imageability and other psycholinguistic and interventional features helps improve performance in both natural and spoken language processing applications

Additionally, we reproduce earlier SOTA approaches and new ideas on a new smaller, balanced and unbiased ADReSS dataset and we believe these deep learning based though not SOTA on ADReSS dataset at the moment, have potential to perform better with more and more data.

And there's plenty of directions for future work, for example all methods right now utilize manually generated transcripts for NLP approaches. Whereas scaling these methods to real world applications would require them to use automated transcription services which is ...

... known to affect disfluency estimates and other features crucial for dementia detection. Moreover, we need better user interfaces in place to scale data collection! Some of these ideas currently underway at @CogneuroBITSGoa

Many thanks to several members at @SforAiDL and @CogneuroBITSGoa for their continued encouragement and support! 🙂