Psychiatrists characterize schizophrenia, a mental condition with devastating effects on those who suffer it, by a set of intuitively understandable concepts including “poverty of speech” and “flight of ideas.” These concepts, however, are subjective in the sense that their quantification depends significantly on the particular training and ultimate judgment of individual psychiatrists. The evaluation of a patient is not only subject to this uncertainty, but also to the availability of these highly trained professionals, and to the clinical facilities where evaluation is usually performed.
The IBM Research team for Computational Psychiatry and Neuroimaging has been developing methods intended to mitigate these limitations using Artificial Intelligence (AI) to provide a consistent (i.e. evaluator-independent) neuropsychiatric evaluation from speech samples, which could be delivered remotely. In collaboration with academic researchers, we published results in 2015 showing that it is possible to use AI to model differences in speech patterns between high-risk patients who subsequently developed psychosis and those who did not. Those results were obtained by analyzing baseline interviews in which patients talk about themselves for close to one hour, using the AI method Natural Language Processing (NLP) to quantify the notion of “poverty of speech” as syntactic complexity, and “flight of ideas” as semantic coherence. While successful, this approach had several limitations: only one cohort, from a single location, and using a single evaluation protocol.
The figure plots the average and standard deviation of the values of the two main factors that provide for conversion (i.e. transition from at-risk to onset) prediction. There are four groups: controls (CTR), patients at-risk (clinical high risk) who do not transition to psychosis (CHR-), patients who do eventually transition (CHR+), and patients who have already transitioned to psychosis (FEP, i.e. first episode). The horizontal axis is related to the maximal value of semantic discourse coherence; observe that the control population have higher values. The vertical axis corresponds to the variance of the coherence; in this case the patients who eventually transition to psychosis (CHR+) show a higher value, meaning that they oscillate between being highly incoherent to being highly coherent. Those who have already converted (FEP), on the other hand, have low maximal coherence, and low variance; i.e. the stay always incoherent.
In a recent paper, Prediction of psychosis across protocols and risk cohorts using automated language analysis, published in the top-ranked psychiatry journal, World Psychiatry, we took steps toward overcoming these limitations. Studying a different and larger patient cohort, from a different location and a very different evaluation protocol (patients being asked about how they understood a story they read), we built a retrospective model of patient speech patterns that would have predicted with 83 percent accuracy whether a patient subsequently developed psychosis. Significantly, this model also would have predicted with 79 percent accuracy whether patients in the earlier cohort would have developed psychosis. That is, the automated, AI-enabled analysis of the speech patterns of these at-risk patients may have the potential to be highly predictive of the future onset of psychosis in a way that bridges differences in recruitment criteria between institutions, and the specifics of the probing protocol.
We believe this is a significant step towards the goal of developing a tool for mental health practitioners, caregivers and patients themselves with a tool that can expand and multiply the reach of neuropsychiatric assessment outside of the clinic. Stay tuned to hear more about the work we are doing in several other mental health conditions such as depression, Parkinson’s and Alzheimer’s diseases and chronic pain, among others.
The recent paper is a collaboration with Mt. Sinai School of Medicine, Columbia University, UCLA and the Universities of Melbourne, Australia, and Buenos Aires, Argentina.
Corcoran, C. M., Carrillo, F., Fernández-Slezak, D., Bedi, G., Klim, C., Javitt, D. C., . . . Cecchi, G. A. (2018). Prediction of psychosis across protocols and risk cohorts using automated language analysis. World Psychiatry, 17(1), 67-75. doi:10.1002/wps.20491
IBM is launching today a broad range of new AI-powered capabilities and services to help CIOs automate various aspects of IT development, infrastructure and operations, including IBM Watson AIOps and Accelerator for Application Modernization with AI. As is the case with much of IBM’s AI development, significant portions of the technologies underlying Watson AIOps and the Accelerator were born out of IBM Research.
The 45th International Conference on Acoustics, Speech, and Signal Processing is taking place virtually from May 4-8. IBM Research AI is pleased to support the conference as a bronze patron and to share our latest research results, described in nine papers that will be presented at the conference.
IBM Research's contributions to CHI 2020 focus on creating and designing AI technologies that center on user needs and societal values, spanning the topics of novel human-AI partnerships, AI UX and design, trusted AI, and AI for accessibility.