/Psychiatric-Disorders-With-Speech

In this project, we wish to identify psychiatric disorders through patient's speech

Primary LanguageJupyter NotebookMIT LicenseMIT

Psychiatric Disorders With Speech

Mail to nikhil.garg@usherbrooke.ca to collabarate

Mental health is important at every stage of life, from childhood and adolescence through adulthood, even more in the digital world created due to the ongoing pandemic and lockdown. Although we have huge medical research on these disorders, there is no lab test to diagnose any of the disorders. Moreover, early detection of degenerative disease can prevent irreversible changes if detected at a later stage during EEG, MRI, etc. In today's world of voice automation, virtual classes, and meetings, speech data is the easiest sample that can be collected from anyone, using existing devices like phones, laptops.

In this project, we wish to identify psychiatric disorders through patient's speech

Review paper : https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7042657/

List of articles of interest : https://tinyurl.com/y6ojfq56

Task 1 : Import the data from Mobile Device Voice Recordings at King's College London (https://zenodo.org/record/2867216#.X5-_tNtS-gQ). Create a PyTorch dataloader for the dataset and layout procedure for adding more datasets in future.

Task 2 : Create scripts to visiualize the data using matplotlib, or any other python package. Implement any Machine Learning algorithm on the loaded dataset to define training and evaluation flow.

Task 3 : Reproduce the results from https://www.ijraset.com/fileserve.php?FID=31054

Task 4 : Implement any Machine Learning algorithm with or without(end-to-end) feature engineering . To start off, use https://github.com/espnet/espnet.

Task 5 : Create pytorch data loader for all the speech datasets listed below. (about 20)

Task 6 : Convert the speech dataset to words for NLP analysis.

We wish to explore state of the art RNN architectures (GRU, LSTM, seq-to-seq, etc.), transfer learning and meta-learning (learning to learn) approaches for this task. These algorithms would be deployed in real time, low power mobile devices for identifying mental disorders from voice-calls, meetings, etc.

Speech-nlp-datasets

Links to publicly available datasets for modeling health outcomes using speech and language :

Speech-based Corpora

TalkBank Project

  • [Corpus] CHILDES Database
    Contains speech of children with different conditions (e.g. Autism, Down's syndrome, hearing impairment) and across different languages (e.g. English, Dutch, Greek, Mandarin).
    MacWhinney, B. (2014). The CHILDES project: Tools for analyzing talk, Volume II: The database. Psychology Press.

  • [Corpus] DementiaBank (from TalkBank)
    Contains recordings of individuals with dementia across different languages. Includes around 400 subjects, most notable in size and containing control subjects is:

    • English Pitt: Longitudinal neuropsychological assessments of 319 subjects (dementia + control) performing Cookie Theft, Word Fluency, Story Recall, and Sentence Construction task. (Becker et al., 1994)
  • [Corpus] Clinical TalkBank
    In addition to DementiaBank, TalkBank contains:

    • RHDBank individuals with Right-Hemisphere Disorder
    • TBIBank individuals with Traumatic Brain Injury
    • AphasiaBank a communication disorder affecting ability to speak, write, and understand language due to some trauma to language parts of the brain.
    • FluencyBank contains individuals with language disfluencies due to being a second language learner, or due to stuttering.

Text-based Corpora

  • [Corpus] MIMIC III (Medical Information Mart for Intensive Care)
    Contains medical details and outcomes of 40,000+ patients (e.g. demographics, vital signs, laboratory tests, medications) as well as 2M+ free-text written medical notes from medical personnel (e.g. physicians, nurses, etc.). (Johnson et al., (2016)).

  • i2b2/UTHealth NLP Task (contact authors for corpus?)
    Contains emergency medical records for 296 patients at Partners HealthCare and medical discharge and correspondance notes between medical personnel. Kumar et al., (2014) describes how the data was processed, and Stubbs et al. (2014) describes the 2014 task of identifying risk factors for heart disease over time.

  • Nun Study (contact authors for corpus?)
    Diaries of 93 nuns to used to evaluate cognitive impairment (Alzheimer's disease) in later life. Also contains neuropsychology tests and autopsy information. Study was authored by (Snowdon et al.,(1996))

Aspects to explore

  1. Acoustic features independent of context. Pitch, pauses, pronounciation etc.
  2. Language features : The patterns present in speech-text converted corpus.