Speech-based Assessment of Relapse Risk in People With Psychosis

Condition(s) studied

Investigational drug(s) / intervention(s)

Weekly online speech assessments via a smartphone app collect speech and self-report data. Recordings are securely transferred and analyzed by an AI-based backend to calculate relapse risk scores.: The intervention is a weekly online assessment of speech, to detect subtle characteristics of psychotic speech. These recordings are made through a speech data collection tool: a smartphone app with implemented tasks for the collection of behavioral response data (speech and self-report). Data is then transferred to a safe repository (TSD) to be analysed by an AI-based speech data analysis algorithm: a backend system will run the predictor code to calculate automatic relapse risk scores.

Study summary

This observational, multinational study assesses the feasibility of speech and self-report data collection across six languages for Artificial Intelligence (AI)-driven relapse risk estimation in psychosis. Over 12 months, patients at risk of relapse and healthy controls will provide weekly speech recordings and self-report data for automated analysis. Risk scores will be stored but not shared with treating clinicians. Independent clinical evaluations ensure data quality and validation. The study lays the foundation for future Clinical Decision Support System (CDSS) research and explores novel speech markers for relapse prediction while minimizing participant burden.

Eligibility

Primary outcome measure(s)

• Feasibility: User adherence — From enrollment to the end of the study at 12 months
  Proportion of users adhering to their allocated tasks. Task completion is automatically assessed for each user. A user is regarded as adherent, if they complete at least 33% of their tasks across all sessions.
• Feasibility: Transcription quality — From enrollment to the end of the study at 12 months
  Proportion of users for which high quality transcripts are produced. Transcription quality is measured by the Word Error Rate (WER) between Automated Speech Recognition (ASR) and human corrected transcripts. A user with an overall WER of at most 35% is regarded as having high quality transcripts
• Feasibility: Usability of recordings — From enrollment to the end of the study at 12 months
  Proportion of recordings that can be used by the AI algorithm. Usability of each record is assessed by HITL reviewers based on the three questions (i) has the user interpreted the question correctly, and (ii) is the audio (at least partially) audible and (iii) comprehensible. The three questions must both be answered with yes for the recording to be usable
• Overall system usability — 6 months and 12 months
  Proportion of users rating the app as usable. Usability is evaluated using the System Usability Scale (SUS), providing a standardized usability score (range 0-100) based on a 10-item questionnaire. We regard a score of at least 70 as usable.
• Performance of the monitoring system for relapse prediction — From enrollment to the end of the study at 12 months
  Assessed by the Area Under the Receiver Operating Characteristic Curve (AUC). Thresholds for relapse prediction will be explored as part of the analysis

Trial sites (7)