AI-Supported Dementia Diagnosis in Low-Resource Settings: Protocol for a Prospective Diagnostic Alignment Study in Accra, Ghana

Dr. David Brodie-Mends

Department of Neurology, Korle-Bu Teaching Hospital | University of Ghana | Atlantic Fellow for Equity in Brain Health at GBHI

david.brodie-mends@gbhi.org

Background

Neurodegenerative disorders, including dementia, are increasing rapidly in low- and middle-income countries (LMICs), where access to specialist diagnostics and disease biomarkers remains limited. In Ghana, diagnosis of these conditions relies primarily on clinical phenomenology due to constrained neuroimaging and laboratory resources. Artificial intelligence (AI)–based clinical decision support systems, developed largely in high-resource settings, have the potential to augment specialist care but may perform unpredictably when deployed in culturally and infrastructurally distinct environments. Evidence evaluating how these AI systems align with clinician judgment in real-world, low-resource dementia care settings is limited.

 

Methods

We will conduct a prospective, single-center observational study at the Adult Neurology Clinic of Korle-Bu Teaching Hospital, Accra, Ghana. Adult patients presenting with suspected dementia syndromes will undergo parallel assessments: (1) a standard neurologist-led evaluation and (2) an AI-supported assessment using a structured minimum data set. An independent expert adjudication panel will establish a consensus clinical diagnosis using baseline data and up to six months of follow-up, serving as a pragmatic reference standard in the absence of universal biomarkers. Diagnostic alignment between clinicians, AI outputs, and adjudicated diagnoses will be quantified using Cohen’s Kappa and Gwet’s AC1 statistics. Cases of discordance will undergo mixed-methods root cause analysis to characterize informational, inferential, and contextual or cultural sources of divergence. Secondary analyses will assess workflow feasibility and performance differences across language and education subgroups.

 

Results

This study will generate quantitative estimates of diagnostic alignment between AI-supported assessments, clinician judgment, and expert consensus diagnoses under resource-constrained conditions. Qualitative analyses will characterize recurrent patterns of divergence, including the impact of missing data, contextual symptom interpretation, and population differences from the AI training domain. Subgroup analyses will explore whether alignment varies by language, education (as a proxy for literacy), or access to biomarkers.

 

Conclusions

By shifting emphasis from accuracy alone to diagnostic alignment and contextual divergence, this study provides a framework for evaluating AI-based clinical decision support in real-world dementia care in LMIC settings. Findings will inform whether and under what conditions AI systems developed in high-resource environments can be responsibly adapted to support equitable neurological care in resource-constrained contexts.