Abstract A017: Personalizing treatment selection for prostate cancer using causal machine learning
Bibliographic record
Abstract
Abstract Background: New treatments for prostate cancer (PC) offer promise but challenge clinicians to optimize sequential selection. Subgroup analyses may identify factors, such as disease volume, that predict treatment benefit, but these analyses overlook the impact of the combined effect of multiple factors. To address this gap, we applied causal machine learning to phase III clinical trials to predict each patient’s individualized treatment effects (ITEs) given their unique multivariable baseline characteristics. Here, we quantify how frequently ITE models can identify significant treatment effect variation in trials of radiotherapy (RT) or systemic interventions for localized or metastatic PC. Methods: We included four PC phase III trials that met the following criteria: in the NCI NCTN Data Archive, >300 participants, and met target enrollment. One trial evaluated RT for localized PC (NCIC-PR.3), two evaluated systemic therapy for localized PC (RTOG-96-01 and RTOG-0521), and one evaluated systemic therapy for metastatic PC (CHAARTED). Trial-specific data included baseline characteristics from Table 1 of the primary publication and the outcome with the shortest median time to event across the trial’s population. Causal Survival Forest, an algorithm to estimate ITE, was iteratively trained in 4/5 of each trial to predict Restricted Mean Survival Time conditional on patient characteristics (cRMST) in the remaining 1/5, resulting in predictions for all trial participants. The median p-value for the Qini coefficient, a metric that quantifies how well the model orders patients by most to least benefit, across five repeats of model training was used to discern statistically significant performance. The variable importance of each model was quantified using Friedman’s H-statistic. Results: Trial-specific ITE models accurately identified those who experienced benefit or harm in 2/4 trials: CHAARTED, a trial evaluating the addition of systemic therapy (docetaxel) to androgen deprivation therapy (ADT) for metastatic hormone-sensitive PC (Qini p-value < 0.05) and NCIC-PR.3, a trial evaluating pelvic RT for localized PC (Qini p-value < 0.05). In CHAARTED, all patients were predicted to benefit from docetaxel, but to varying degrees (cRMST range 0.19 to 1.42 years). The variables most predictive of treatment response were time from ADT treatment to randomization, baseline PSA, age, and volume of metastatic disease. In NCIC-PR.3, most patients benefited from pelvic irradiation, but some experienced harm (cRMST range -0.05 to 0.36 years). The variables most predictive of treatment response were age, PSA < 20 or > 50, Gleason score > 8, and prior hormone therapy. Conclusions: With clinical characteristics alone, ITE models accurately predicted personalized treatment effect estimates in 2/4 trials. Once validated, ITE modeling could help clinicians optimize treatment selection in PC. This work was supported by the NLM (5T15LM007359). Data was shared through NCI’s NCTN Data Archive, with approval from sponsors: NCI, ECOG, SWOG, NRG Oncology, RTOG, MRC, and the NCIC Clinical Trials Group. Citation Format: Emma Graham Linck MS, Alex Spicer MS, Marina Sharifi MD PhD, Guanhua Chen PhD, Mark Craven PhD, Nataliya Uboha MD PhD, Mark Burkard MD PhD, Matthew Churpek MD MPH PhD. Personalizing treatment selection for prostate cancer using causal machine learning [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr A017.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.046 | 0.129 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.002 | 0.003 |
| Bibliometrics | 0.002 | 0.002 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.002 | 0.002 |
| Open science | 0.002 | 0.002 |
| Research integrity | 0.002 | 0.003 |
| Insufficient payload (model declined to judge) | 0.006 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".