Pathology

SPARK: agentic AI generates biologically meaningful histopathology concepts across cancers

SPARK uses language-guided agents to convert biological concepts into analyzers of raw histopathology without additional model training. The framework was evaluated across 18 cohorts and over 5,400 patients spanning five cancer types, including a 625-patient breast cancer spatial biology set. SPARK-derived concepts correlated with prognosis, known pathological variables, and predictive biomarkers across prognostic and predictive tasks. A human-interaction module allows expert oversight, but the authors state prospective clinical validation is still required. 1

Expert regrading reveals misclassification drives higher CIN3+ risk during CIN2 surveillance

In 437 women aged 23–40 under CIN2 active surveillance, three experts regraded initial biopsies and upgraded 12.8% to CIN3. Overall, 39.6% had a CIN3+ event over a median follow-up of 28 months, with risk highest after expert CIN3 classification. Adjusted analyses showed expert CIN3 diagnosis associated with increased CIN3+ risk (aRR 1.37), especially in older women and those with high-grade cytology. These results imply diagnostic variability substantially affects progression estimates during conservative management. 2

EMR-integrated risk assessment increases identification of patients meeting hereditary cancer counseling criteria

A single-center nonrandomized trial invited 1,685 adults to complete an EMR-integrated family and personal history risk tool; 544 received a risk report. Among completers, 28.5% (155/544) met guideline criteria for genetic counseling, but only 31 attended counseling. The platform exposed gaps in usual care identification methods and could triage patients before visits. Limitations include single-arm design and modest downstream counseling uptake, limiting conclusions about clinical impact. 3