A rules-based AI triage system to identify NCCN-indicated genetic testing in prostate cancer: A performance analysis.

Abstract

54 Background: Somatic and germline testing in prostate cancer can inform overall prognosis, treatment sequencing decisions, and familial risk through cascade testing. Despite clear guidelines from organizations like the AUA, ASCO and NCCN, a significant gap remains between recommendations and clinical practice, with two-thirds of men in the US with metastatic prostate cancer not receiving consensus guideline compliant testing. This study evaluates a rules-based artificial intelligence (AI) system that identifies patients who meet NCCN criteria for somatic and germline testing, with the end goal of improving adherence to the clinical guideline. Methods: We developed a rules-based AI (ChatGPT 5.0) system to mirror NCCN v1.2023 guidelines on somatic and germline testing for prostate cancer, producing three actionable states: "indicated," "consider," and "not indicated.” Structured inputs included age at diagnosis, TMN stage, current risk group, grade, and nodal/metastatic status. Somatic rules: indicated if M1; consider if castration-resistant prostate cancer (CRPC) without M1, N1, or localized high/very high-risk disease (T3a–T4, Grade Group ≥4, or PSA >20); otherwise not indicated. Germline rules: indicated if M1, N1, or localized high/very high risk; consider if intermediate risk with age ≤60 or CRPC without metastasis; otherwise not indicated. Metrics were assessed in a retrospective cohort of 259 prostate cancer patient charts seen at Oncology Hematology Care in 2023. Gold standard was manual review of patient charts. Results: Average age at diagnosis was 68 years and 54% of patients had stage IV disease. The AI system demonstrated high sensitivity and specificity. Germline false negatives (n=7) represented cases where testing was considered based on manual review of non-structured data, including family history. Conclusions: This rules-based AI demonstrated excellent performance in identifying NCCN-indicated genetic testing within a real-world cohort of 259 charts, achieving perfect accuracy for somatic testing and near-perfect accuracy for germline testing. Limitations include the retrospective, single-practice design and reliance on structured inputs. Overall, this approach is readily implementable and may improve guideline adherence; multi-site validation and EMR integration are planned. Accuracy of rules-based triage for NCCN-indicated genetic testing. TP FP FN TN Sensitivity (95% CI) Specificity (95% CI) Somatic 195 0 0 64 100% (98–100) 100% (94–100) Germline 208 0 7 42 96.7% (93–98) 100% (92–100)

Ähnliche Arbeiten