Artificial intelligence (AI)-powered diagnostic support for stroke via Telegram bot: preliminary findings

BACKGROUND: Annually, stroke affects over 15 million people globally. Early intervention is critical in the management of stroke. However, the "golden hour" opportunity for timely intervention is often missed in underserved regions where experts in neuroimaging are scarce, with some countries having 1 per million inhabitants in comparison to high-income countries with 10 per million inhabitants. This study explores the feasibility of a lightweight AI-powered diagnostic support system for stroke deployed via a widely accessible platform (Telegram) to help clinicians in underserved settings. METHODS: This study utilized a comparative analysis of two transfer learning models (EfficientNetB0 and MobileNetV2) which were employed for training, using 6,650 publicly available, anonymized but radiologist annotated brain CT images from three classes: normal (n = 4,427), hemorrhagic stroke (n = 1,093), and ischemic stroke (n = 1,130). An 80/20 split was executed at slice-level as a result of lack of patient identifiers which this study acknowledges as a limitation. A separately sourced Kaggle dataset with labels but no metadata on annotator provenance was used for quasi-external validation with further evaluation using standard performance metrics. The selected model was integrated into a Telegram bot (@BrainfloBot) with real-time inference capabilities, and automatic data deletion to ensure privacy compliance. RESULTS: EfficientNetB0 demonstrated superior performance over MobileNetV2, achieving 99% accuracy (95% Cl: 98.2-99.5%) with excellent inter-rater reliability (κ = 0.985) during training and internal validation with precision and recall values exceeding 96% across all classes. While quasi-external validation using the balanced set, the selected model achieved a robust performance with 95% accuracy level (95% CI: 90.3%-97.9%, n = 150) with excellent agreement (κ = 0.925). For the unbalanced set, 97% accuracy level (95% Cl: 95.77%-98.03%) with excellent agreement(κ = 0.948) The model showed particularly strong performance in hemorrhagic stroke detection (98% precision, 96% recall), critical for preventing inappropriate thrombolytic therapy. CONCLUSIONS: These preliminary findings demonstrates a successful integration of a novel AI-powered stroke diagnostics with a widely accessible messaging platform. This system can potentially close neuroimaging gaps in underserved regions where neuroimaging specialists are scarce. However, the study limitations include lack of patient-level separation, annotation provenance for external validation dataset, possible selection bias and lack of real world testing. Future work can integrate real world validation, out of distribution handling and robust testing with clinical images.