Artificial Intelligence (AI) and Risk Assessment in Metastatic Hormone-Sensitive Prostate Cancer.

Risk assessment in metastatic hormone-sensitive prostate cancer (mHSPC) requires clinicians to integrate information from multiple sources. CT, MRI, bone scans, and prostate pathology reports may need to be reviewed together to establish TNM stage, identify and count metastatic lesions, determine whether visceral disease is present, and characterize tumor burden. These findings can then be mapped to established risk frameworks, including CHAARTED, LATITUDE, and, in Taiwan, the Taiwan National Health Insurance (TwNHI) criteria.[1][2][3]

Three frameworks, different questions
CHAARTED: metastatic disease volume. The CHAARTED framework defines high-volume disease as the presence of visceral metastases and/or at least four bone metastases, with at least one bone lesion outside the vertebral column or pelvis. The original CHAARTED trial evaluated adding docetaxel to androgen-deprivation therapy (ADT) in mHSPC and showed an overall-survival benefit with combination treatment.[1]

LATITUDE: high-risk disease. The LATITUDE framework defines high-risk disease when at least two of three features are present: a Gleason score of 8 or higher, at least three bone metastases, or measurable visceral metastases. The criteria were developed in the LATITUDE trial, which evaluated abiraterone acetate plus prednisone in combination with ADT in men with newly diagnosed, metastatic, castration-sensitive prostate cancer.[2]

TwNHI: a reimbursement-based framework. The TwNHI criteria were developed within Taiwan's National Health Insurance system and use a combination of pathologic and metastatic features to define high-risk disease for reimbursement purposes. In the AI study described below, TwNHI high risk required at least two of three features: a Gleason score of 8 or higher, four or more bone metastases with at least one outside the vertebral bodies and pelvis, or measurable visceral metastasis.[3]

The three frameworks overlap, but they are not interchangeable. CHAARTED primarily characterizes metastatic disease volume, LATITUDE identifies a high-risk clinical phenotype, and TwNHI incorporates criteria used within Taiwan's health-insurance framework.[1][2][3]

What the clinical task requires

Applying these criteria to a patient's record involves more than locating a single data point. The information must first be retrieved from multiple reports and then integrated

What has been tested with large language models?

A 2025 study published in World Journal of Men's Health evaluated four general-purpose large language models—ChatGPT-4-turbo, Claude-3-opus, Gemini-Pro 1.0, and ChatGPT-3.5-turbo—on information-retrieval and risk-assessment tasks using simulated multimodality work-up reports from 350 patients with stage IV prostate cancer. The reports incorporated CT, MRI, bone-scan, and pathology information. The study evaluated three risk-assessment tasks: LATITUDE high risk, CHAARTED high volume, and TwNHI high risk.[3]

The investigators compared model outputs with a consensus assessment by experienced urologists. Using the study's ensemble-voting approach, ChatGPT-4-turbo achieved accuracies of 90.1% for LATITUDE, 90.7% for CHAARTED, and 91.6% for TwNHI. Information-retrieval performance for TNM staging was also high, with reported accuracies ranging from 87.4% to 94.2% across the models and tasks evaluated.[3]

These findings are notable because the models were not simply asked to retrieve a single fact. They were required to extract information from multiple clinical reports and apply several predefined rules to reach a risk classification. The TwNHI task is also distinct because the criteria are specific to Taiwan's health-insurance system.[3]

What the study does—and does not—show

The results demonstrate that general-purpose LLMs can perform structured information retrieval and rule-based risk classification from simulated prostate cancer work-up reports under the conditions tested. They do not establish that an LLM can independently perform clinical risk assessment or safely determine treatment for individual patients.

The study used simulated reports rather than prospectively collected clinical records, and the investigators specifically noted the potential for misinterpretation to affect clinical decision-making. The authors concluded that further validation is required in other cancers and clinical settings.[3]

For AI in oncology, this distinction truly matters. Extracting the right facts and correctly applying a defined classification rule is a measurable task. Whether an AI system can reliably integrate those classifications into real-world clinical decision-making is a separate question.

References
[1] Sweeney CJ, Chen YH, Carducci M, Liu G, Jarrard DF, Eisenberger M, Wong YN, Hahn N, Kohli M, Cooney MM, Dreicer R, Vogelzang NJ, Picus J, Shevrin D, Hussain M, Garcia JA, DiPaola RS. Chemohormonal Therapy in Metastatic Hormone-Sensitive Prostate Cancer. N Engl J Med. 2015 Aug 20;373(8):737-46. doi: 10.1056/NEJMoa1503747. Epub 2015 Aug 5. PMID: 26244877; PMCID: PMC4562797.
[2] Fizazi K, Tran N, Fein L, Matsubara N, Rodriguez-Antolin A, Alekseev BY, Özgüroğlu M, Ye D, Feyerabend S, Protheroe A, De Porre P, Kheoh T, Park YC, Todd MB, Chi KN; LATITUDE Investigators. Abiraterone plus Prednisone in Metastatic, Castration-Sensitive Prostate Cancer. N Engl J Med. 2017 Jul 27;377(4):352-360. doi: 10.1056/NEJMoa1704174. Epub 2017 Jun 4. PMID: 28578607.
[3] Yuan LH, Huang SW, Chou D, Tsai CY. The In-depth Comparative Analysis of Four Large Language AI Models for Risk Assessment and Information Retrieval from Multi-Modality Prostate Cancer Work-up Reports. World J Mens Health. 2025 Oct;43(4):918-933. doi: 10.5534/wjmh.240173. Epub 2024 Dec 2. PMID: 39743220; PMCID: PMC12505481.

Featured image: © 2926 Ch LLM and AI in Oncology. Used with permission.
DOI