Abstract B001: Jarvais: A modular framework to standardize machine learning workflows and accelerate reproducible AI in oncology – benchmarking against a human-developed model for predicting emergency department visits during cancer treatment
Bibliographic record
Abstract
Abstract Background: Artificial intelligence (AI) continues to advance oncology research, yet inconsistent development pipelines impair reproducibility and introduce bias. A lack of transparency in AI research1—including undisclosed preprocessing, omitted hyperparameters, and unshared code—undermines validation and trust. To address these issues, we present JARVAIS (Just A Really Versatile AI Service), an open-source Python package that standardizes machine learning (ML) workflows for oncology, improving reproducibility and mitigating bias through modular automation of data preprocessing, model training, and interpretability. Methods: JARVAIS (pmcdi.github.io/jarvais/) is composed of three core interoperable modules: (1) Analyzer for data quality audits and bias detection, (2) Trainer for automated feature selection, model training, and hyperparameter optimization, and (3) Explainer for interpretability and post-hoc fairness audits. We assessed JARVAIS by comparing it to AIM2REDUCE-ED (Emergency Department), a manually developed model predicting 30-day ED visits among gastrointestinal cancer patients on systemic therapy at Princess Margaret Cancer Centre, currently in silent deployment with strong performance2,3. JARVAIS was used to generate a model predicting the same outcome from the longitudinal retrospective de-identified Electronic Health Record data used to develop the AIM2REDUCE-ED model. Results: The Analyzer module in JARVAIS automatically handled missing value imputation and outlier trimming (0.01–0.99 quantile) and the Trainer module used 5-fold cross validation, consistent with prior manual strategies. The JARVAIS model achieved an AUROC=0.70 and an AUPRC=0.23 on a held-out test set, closely aligning with the original manually engineered model (AUROC=0.69, AUPRC=0.20). In the prospective cohort, the JARVAIS developed model obtained an (AUROC=0.67,AUPRC=0.11) as compared to AIM2REDUCE-ED model (AUROC=0.69,AUPRC=0.16). The Explainer module also flagged patients on treatment regimens (CISPFU+TRAS, CISPCAPE+TRAS) for potential performance bias. These groups were previously identified in manual bias analysis and had worse performance in ED prediction with AUROC=0.27. Despite matching predictive accuracy, JARVAIS achieved results in under 30 minutes with minimal manual effort—automating data processing, model selection, hyperparameter tuning, validation and bias analysis. Conclusion: JARVAIS demonstrates strong potential as a scalable, reproducible platform for rapid ML development in oncology. Benchmarked in a real-world clinical setting, it matched the performance of a human-developed model while accelerating development and ensuring consistency. By embedding explainability and fairness into its workflow, JARVAIS supports clinician understanding of model outputs and fosters trust in downstream use, enabling more transparent and equitable AI integration in clinical practice. References: 1. Haibe-Kains, B. et al. Nature 586, E14–E16 (2020). 2. Grant, R. et al. JCO 41, 1557-1557 (2023). 3. Kabir, M. et al. JCO Oncol Pract 20, 407-407(2024). Citation Format: Joshua Siraj, Muammar Kabir, Sejin Kim, Jiang Chen. He, Baijiang Yuan, Wayne Uy, Tirth Patel, Benjamin Grant, Sharon Narine, Monika Krzyzanowska, Tran Truong, Geoffrey Liu, Clare McElcheran, Robert Grant, Mattea Welch. A modular framework to standardize machine learning workflows and accelerate reproducible AI in oncology – benchmarking against a human-developed model for predicting emergency department visits during cancer treatment [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 B001.
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 | 0.010 | 0.032 |
| Meta-epidemiology (narrow) | 0.003 | 0.002 |
| Meta-epidemiology (broad) | 0.001 | 0.002 |
| Bibliometrics | 0.002 | 0.001 |
| Science and technology studies | 0.001 | 0.002 |
| Scholarly communication | 0.003 | 0.004 |
| Open science | 0.005 | 0.006 |
| Research integrity | 0.001 | 0.004 |
| Insufficient payload (model declined to judge) | 0.020 | 0.018 |
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".