OncoHelper AI — Longitudinal Modeling of Pediatric Cancer Trajectories

OncoHelper AI is developing a predictive framework for pediatric oncology based on the longitudinal trajectory of the patient. The first research question is whether changes in disease state and treatment response over time provide predictive information beyond a conventional static clinical snapshot.

The first experiment compares two models in the same patient cohort, with the same outcome and test patients.

Model A — Static Baseline: baseline clinical, disease-risk and molecular/genomic information available at a defined prediction point.

Model B — Longitudinal: the same baseline information plus available time-indexed data such as treatment phase, sequential disease measurements, early response, MRD or other response markers, laboratory trends, toxicity and progression-related events.

Prototype 0.1 uses retrospective validation. For each test patient, all post-cutoff information is hidden. Both models predict from data available at that time; the later historical outcome is then revealed.

The primary endpoint, prediction horizon and evaluation metric will be fixed before the final test set is evaluated. Model A and Model B will use identical patient-level partitions to prevent an artificial comparison.

Performance will be evaluated with AUROC, AUPRC, calibration and Brier score where appropriate. The primary comparison is the paired Model B minus Model A difference with 95% confidence intervals.

A positive Prototype 0.1 result requires: a prespecified primary performance improvement for Model B whose paired 95% confidence interval excludes zero on held-out patients, without presenting improved discrimination as success if calibration materially deteriorates.

The OncoHelper data structure represents a patient as:

This is the common data foundation for later short- and long-horizon models.

Prototype 0.1: patient history → future clinical outcome.

Prototype 0.2: patient history + actual intervention → subsequent patient state.

Prototype 0.3: patient state + alternative intervention scenarios → alternative future trajectories using causal and counterfactual methods.

The longer-term architecture includes INTERCEPTOR for long-horizon trajectory modeling, TACTICIAN for short-horizon response prediction, SENTRY for trajectory surveillance, and META-AI for evidence quality, model agreement and uncertainty.

The Patient Trajectory Schema and Model A versus Model B design are defined. TARGET ALL and TARGET Neuroblastoma have been audited; both are useful for architecture and baseline development but lack sufficient treatment-response depth for the primary experiment. A coherent pediatric cohort with baseline, treatment, measured response and later outcome is now being selected.

The current development target is the first working prototype and initial retrospective validation results by mid-September 2026. No model-performance result is assumed in advance.