Artificial intelligence improves prediction of long-term outcomes

An artificial intelligence (AI) tool, trained on data from 5 randomised-controlled trials, proved to be more effective than clinical prognostic markers for predicting long-term outcomes in patients with prostate cancer.

To risk stratify patients, most guidelines, including the NCCN guidelines, use parameters like prostate-specific antigen (PSA), digital rectal exam (DRE), and Gleason score. However, this risk stratification tool is suboptimal in terms of performance. More recently, tissue assays have been developed assessing genomic biomarkers, but they are expensive and slow. AI tools leveraging digital pathology may provide solutions to these problems because they are quick, robust, and can be developed on thousands of patient samples without consuming any tissue. Dr Osama Mohamad (University of California San Francisco, CA, USA) presented a prognostic AI tool that was trained and validated using data from 5 phase 3 randomised trials, by leveraging multi-modal deep learning on digital histopathology [1].

Histopathology image data was generated from pre-treatment biopsy slides in 5 NRG Oncology radiotherapy prostate cancer trials (RTOG 9202, RTOG 9408, RTOG 9413, RTOG 9910, and RTOG 0126). Of 5,654 eligible patients, 16,204 digitalised histopathology slides of pre-treatment biopsy samples were randomly split into training (80%) and validation (20%) cohorts. A multi-modal artificial intelligence (MMAI) architecture was developed to take clinicopathologic and image-based data as input and predict binary outcomes.

After training, locking, and evaluating on the validation cohort, the MMAI prognostic model showed superior discrimination in comparison with the NCCN model (based on PSA, T-stage, and Gleason score) for 5- and 10-year distant metastases, 5- and 10-year biochemical failure, 10-year prostate cancer-specific survival, and 10-year overall survival (see Figure).

Figure: Evaluation of incremental benefit of various data [1]



*NCCN = Gleason combined + baseline PSA + T-stage; **Pathology images + NCCN + Gleason primary + Gleason secondary + age; BF, biochemical failure; DM, distant metastases; OS, overall survival; PCSS, prostate cancer-specific survival

Based on these results, Dr Mohamad concluded that the AI tool can successfully predict long-term, clinically relevant outcomes for patients with prostate cancer.

1. Esteva A, et al. Development and validation of a prognostic AI biomarker using multi-modal deep learning with digital histopathology in localized prostate cancer on NRG Oncology phase III clinical trials. Abstract 222, ASCO GU 2022, 17–19 February.

 

Copyright ©2022 Medicom Medical Publishers



Posted on 8 April 202217 May 2024