MediDocker: An AI-Enabled Healthcare Intelligence Ecosystem Integrating Clinical Decision Support, Predictive Analytics, and Emergency Response

Patients keep their health information in places that do not talk to each other: handwritten prescriptions, PDF blood reports, appointment cards and paper bills, and in an emergency nobody can quickly see their blood group or allergies. This paper presents MediDocker, a web platform that keeps these records in one secure place and explains them in plain language. A vision-language model (Qwen2.5-VL, run locally) only copies what a prescription says; every medicine is then verified against RxNorm, an Indian brand dictionary and a hand-checked drug table, and its purpose and warnings come from verified sources, never from the model. Blood (CBC) reports are encrypted, read with a PDF text layer or PaddleOCR, checked by seven plausibility rules, and followed across reports with reference change values, a personal set point and Theil–Sen projection. A LightGBM model trained on 14,134 laboratory-confirmed NHANES participants screens for iron deficiency from the CBC alone and answers only when it is confident, otherwise advising a serum ferritin test. On 1,945 unseen participants from 2021 onwards it reached an AUROC of 0.836 (95% CI 0.816–0.857) and 94.2% accuracy on the 10.6% of cases it decides, with sensitivity 99.6% and specificity 99.3% at its two cut-offs. A record-aware assistant detects emergencies with rules before any language model is used, and an Ed25519-signed emergency QR code can be verified offline by anyone, without login