Abstract
Background Rich clinical information is often stored in handwritten patient charts, but there are limited systems for converting these routine records into usable data for research, care, service improvement, and policy. This process paper describes the development of Hekima, a locally hosted, AI-assisted mental health data platform that converts paper psychiatric records into structured longitudinal datasets for research. Methods We reviewed implementation documents generated during the Digitising Psychosis Project of the Mental Health Data Prize Africa program. We reconstructed the development process, major adaptations, technical and governance decisions, early analytic outputs, capacity-building activities, public communication, and scale-up pathway. Results Hekima evolved from a psychosis chart digitisation project into a broader mental health data platform. Hekima followed a “one pipeline, two outputs” model. Each scanned document was transcribed, reviewed, redacted, and structured into a longitudinal patient record, but also prepared for patient matching with existing electronic medical record systems. The project digitised and archived 203,020 images from 7,164 files. Early OCR-based extraction demonstrated feasibility but performed poorly on handwritten psychiatric notes. This prompted a redesign toward locally hosted vision-language models, redaction of identifiable information, extraction, and mandatory human review. Lived-experience partners contributed to privacy safeguards, engagement interpretation, data-use acceptability, patient-rights protection, public communication, and independent process review. Conclusion Hekima demonstrated that handwritten psychiatric records in a low-resource setting could be transformed into usable mental health data infrastructure but required data governance, lived-experience oversight, capacity building, communication, and engagement with policymakers.