Skip to navigation Skip to main content Skip to footer

Approved Research

Multimodal Deep learning and Big Data on Distinguishing Disorders and Their Social, Individual and Biological Risk Factors

Principal Investigator: Dr Chao-Gan Yan
Approved Research ID: 88866
Approval date: December 22nd 2022

Lay summary

Our research mainly aims at distinguishing mental disorders from each other. Given to current clinical practice, the diagnosis system of mental disorder mainly relies on the experts' oral inquiry. A consistent and subjective evaluation from deep learning based on big data would enhance the credibility of diagnosis and improve the efficiency in medical practices. We wish to reduce the rate of misdiagnosis and accelerate precision medicine by means of creating an intelligent diagnosis system, which can give preliminary classification of mental state of each patient based on tens of thousands of precedents.

The diagnosis, treatment and prognosis, of mental disorders have been suffering from comorbidity for a long time. The manifested symptoms from different disorders could share a lot of similarities and get same diagnosis and medicine cure, which leads to unnecessary financial and mental burdens for patients. As neuroscience steps into big-data era, we have more access to data resources where we could investigate the underlying mechanisms of mental disorders from multi-dimensions. In artificial intelligence field, pictures, words and audios are pretrained unimodally and combined in embedding layer to activate the ability to transform reciprocally. Hence here we intend to take the advantage of different modalities and their dependent/independent relationships to construct a multimodal neural network which can encompass the features from brain imaging perspectives of psychiatric disease from thousands of samples, and interpret our discovery from biological, life style and social aspects. Recently Nature reports that for brain-wise association studies, sample size not more than 1,000 would crash its credibility for poor reproducibility and inflated effect size. Revelation of such phenomenon actually advocates the usage of big dataset and corresponding progress on acquisition and processing perspective to dig into the plentiful treasure underneath sea of data. Our proposal of being accessible towards a wide range of phenotypes and neuroimaging is keeping pace with such scientific vision.

The project lasts at least 18 months. The prominent contribution of our study is to provide applicable classifier to provide clinical-level preliminary diagnosis given to multimodal brain imaging of each participant. Besides, we will dig into the genetic level, behavioral level and social level to explore which factors are highly associated with the distinguishment among psychotic disorders.