PROSPECT: Predicting psychosis outcomes from speech and brain connectivity

Predicting individual trajectories for patients with psychotic illnesses is both a clinical and patient priority. These conditions typically emerge in early adulthood and involve severe symptoms such as hallucinations and delusions. Approximately 25% of patients develop schizophrenia or other long-term, disabling illnesses, whilst others make a full recovery. However, there are currently no reliable methods to predict which young people with psychotic symptoms will experience which of these diverse outcomes. For patients who recover, there is also a lack of tools to monitor for symptom recurrences and predict future relapses ahead of time. Such tools could create a window of opportunity for patients and clinicians to act before symptoms get worse.
New Artificial Intelligence (AI) techniques have significant potential to tackle these challenges. We know that altered language use is one of the main symptoms patients might experience in psychotic illnesses, so brain scans and short recordings of speech stand out as particularly promising data for AI to draw on. I have already led work using AI methods to search for brain connectivity and speech patterns characteristic of psychotic illnesses. I also recently developed new methods to measure patients’ brain connectivity more reliably, and to map the content of patients’ speech in finer detail.
This Fellowship will build on this work, to engineer innovative tools to predict individual trajectories for patients with psychotic illnesses.
Key aims include:

Build the next generation of AI tools to predict likely illness outcomes for early stage psychosis patients, using both brain connectivity and speech information.
Develop AI tools to predict relapse for patients tapering off medication, from speech recordings.
Explore ways in which these tools can be translated to the clinic, by working closely with key stakeholders, including patients and clinicians.

To that end, I will combine recent advances in AI, such as generative AI, with the new methods to measure brain connectivity and study speech data developed by my research group. I will also collect new speech data from patients tapering off medication and therefore at increased risk of relapse, enabling me to engineer tools to predict relapse for this important group. Finally, I will create the first tools to predict how early stage psychosis patients’ symptoms are most likely to change over time from both brain connectivity and speech recordings.
It is crucial that the tools developed are trustworthy, and I will tackle that challenge both technically, for example by investigating new ways to communicate prediction uncertainty to clinicians, and by bringing together a wide range of different stakeholders, including AI experts, clinicians and people with lived experience of psychotic illnesses. People with lived experience have already helped guide the work, for example by suggesting the goal of predicting relapse for patients tapering off medication.
The resulting tools could be used by clinicians to predict which patients are at highest-risk of developing severe, chronic conditions such as treatment-resistant schizophrenia, helping them to prescribe the most appropriate medication at an earlier stage. I also envisage opportunities for patients to use the tools developed here to self-monitor their speech at home and spot early warning signs of relapse. Longer-term, having access to tools to predict outcomes for patients with psychosis is expected to help researchers understand what causes some people to recover more fully than others, and spur future research into new treatments.