An interpretable artificial intelligence model for real-time leukemia screening via routine blood tests across multicenter cohorts

Large-scale screening for leukemia remains challenging due to the absence of simple, scalable, and intelligent tools. Here, we present LeukoAlert, an artificial intelligence framework that transforms routine complete blood count (CBC) data (72 features) into a real-time opportunistic screening and risk flagging tool. Trained and validated on 446,558 records from 203,284 individuals across seven independent centers, the model achieved high discrimination between leukemia and non-leukemia cases (area under the curve (AUC) up to 0.996), with moderate performance for acute versus chronic subtype classification. In a prospective real‑world evaluation of 58,481 unselected records, it maintained robust accuracy (AUC = 0.969), identifying occult leukemia and early relapse in non‑hematology departments. Nine interpretable features grounded in leukemia pathophysiology were identified. By combining the universal availability of CBC tests with an embedded AI approach, LeukoAlert provides a potentially scalable adjunctive tool that may be integrated into compatible laboratory workflows, supporting further evaluation as a risk-flagging tool in routine CBC testing.