Kruthikha
Vishali K

The World Health Organization recommends screening every TB patient for diabetes, but testing capacity is scarce. We built a screening model that ranks TB notifications by recorded diabetes and treatment-era mortality risk using only what the clinic holds on day one of notification — no later-workup fields, no leakage.

Starting from Brazil's SINAN-TB national registry (964,099 records), a field-by-field leakage audit under a single prediction-time rule removed 15 fields recorded during or after treatment (8 treatment-era, 7 downstream), 1,849 post-death notifications, and the 2001–2006 era (30× death under-recording). Analysis cohort: 684,694 notifications, 88 notification-time features. We compared a heterogeneous GNN against XGBoost, logistic regression and MLP on identical inputs and group-aware splits.

Honest negative result: XGBoost reached lock-box AUROC 0.804 (death) / 0.808 (recorded diabetes); the heterogeneous GNN did not beat a plain MLP after equal-budget tuning (0.8156 vs 0.8154 on death; 0.8105 vs 0.8107 on recorded diabetes), consistent with the graph's 101-hub structure sending nearly identical messages across patients. Reported as a finding, not a failure. Isotonic recalibration (fitted on validation) dropped expected calibration error from 0.303 to 0.001, and a WHO-grounded guideline layer attaches verbatim passages to each flagged score (retrieval-grounded explanation without text generation — a deliberate clinical-safety choice). Indian hospital-data retraining planned for the 8th-semester capstone.

Indian commercial invoices are extraction-hard for two reasons at once: a long-tail schema of 174 fields where fewer than 7% are filled per invoice, and natural visual noise that generic vision-language models were never trained for. Off-the-shelf VLMs hallucinate values or produce unparseable output.

RECIQ tackles both problems on Qwen2-VL-7B-Instruct through three core ideas: tokenizer surgery that injects all 174 schema field names as atomic vocabulary tokens, LoRA fine-tuning that concentrates gradient signals on schema-relevant positions, and a grammar-constrained beam decoder that guarantees syntactically valid JSON at every generation step. A downstream GraphSAGE anomaly detector built on RECIQ's structured output reaches 68.1% macro F1 for financial auditing.