LUMANA Field Validation Dataset, Sokoto 2026.

These are datasets generated during the LUMANA field validation study in Sokoto State, Nigeria. The study evaluated an AI-assisted mental health screening system across four complementary areas: trust and acceptability, AI output safety, culturally and linguistically appropriate communication of distress and integration of AI into clinical workflow.This contains three survey-based evaluation datasets and one qualitative dataset:1. Trust and Acceptability ReflectionThis survey dataset captures participants’ perceptions of trust, acceptability, ethical boundaries and appropriate human involvement in AI-assisted mental health screening. Participants reported circumstances in which they would trust or distrust the system, situations in which they would stop using or reject it, and conditions under which human involvement should be required.2. AI Output Safety AssessmentThis survey dataset contains structured evaluations of AI-generated mental health responses presented within eight standardized hypothetical mental health scenarios. Clinicians and lived experience participants assessed the safety, acceptability, appropriateness, potential risks, and suitability of each AI-generated response.The survey captured judgments about whether an output was safe, whether human oversight or escalation was required, the potential level of risk associated with an inappropriate response, and whether the output was suitable for deployment. Reviewers also recorded concerns, emotional responses where applicable.The dataset includes pseudonymized reviewer identifiers, reviewer categories, scenario and AI-output identifiers, safety and deployment classifications, risk assessments, human-oversight and escalation assessments.3. Human Language Quality AssessmentThis qualitative dataset contains Hausa idioms of distress identified and explored during the LUMANA field validation study. It was developed to document culturally meaningful expressions used to communicate psychological or emotional distress and to examine their meaning, context, usage, cultural significance, and potential implications for AI-assisted mental health screening.The dataset includes Hausa expressions or idioms, English translations, contextual or intended meanings, situations and contexts of use, information about typical usage, cultural considerations, and qualitative observations concerning potential AI misinterpretation.4. AI Workflow EvaluationThis survey dataset contains clinician evaluations of the proposed human–AI clinical workflow for the LUMANA system. Five clinicians reviewed eight standardized scenarios (HC-01 to HC-08), with each scenario describing a proposed sequence of AI and human actions for a mental health presentation or clinical situation.Clinicians assessed whether the proposed workflow reflected clinical practice, whether responsibilities were appropriately allocated between AI and human providers, and whether individual workflow steps were safe to automate. The survey also examined potential risks, requirements for human review or escalation, notification preferences, comfort with routine use, trust in the workflow and implementation considerations.The dataset includes pseudonymized clinician identifiers, scenario identifiers, workflow characteristics, structured assessments of decision-making responsibility, automation suitability and risk, escalation and notification preferences and clinician comfort.