Towards integral and holistic validation of lab results

Clinical laboratories rely on simple “if-then” rules to decide which test results need a human double-check: for example, “hemoglobin < 5 mmol/L” or “creatinine change >20 %”. Each value is judged on its own, so the software raises many false alarms and misses odd combinations of values that can signal real danger. By training a neural network on millions of historic test panels, we can let the model learn how laboratory values normally vary together in real patients. Once trained, it should be able to spot result sets that look statistically impossible and highlight them for expert review, shrinking worklists and catching issues rules never see. In this project you will build and evaluate that model, compare it with today’s rule-based system, and quantify the potential gains in both safety and lab efficiency.