Clinical Trial Modeling

A clinical trial is described through several kinds of evidence: its treatment, the disease being studied, the participants it accepts and the way the study is organized. These clues may help predict whether a trial succeeds, how long it runs or which safety problems occur.

The task asks one shared model to answer eleven kinds of question. The scientific idea is that learning about one aspect of a trial might help with another. But sharing can also cause interference: the information useful for predicting approval may differ from what explains dropout.

The data combine TrialBench with three sets of trial-outcome questions. There are roughly 244,000 training records and 61,000 held-out records. A record is a question or target, not a patient; one trial can contribute several records.

Related records must stay in compatible data partitions. Otherwise the model could learn a trial’s details from one task and appear to predict it independently in another. Trial identifiers align records but are not scientific evidence for an answer. The interactive examples use fictional trials to make this distinction visible.

The reference model reads a treatment’s molecular graph, disease codes, numerical fields and trial text through different encoders. A molecular encoder passes messages between bonded atoms; a text encoder represents descriptions and eligibility criteria. Learned layers combine these views into a shared description of the trial.

Different output heads then turn that description into class probabilities, numerical estimates or eligibility text. Missing inputs receive an explicit missingness representation, so an absent measurement is distinguishable from a measured zero. The text encoder is pretrained and fixed; the shared fusion and output components learn from the trial tasks.

Eligibility has far more examples than some outcome questions. Training therefore revisits smaller tasks more often instead of letting the largest dataset dominate every update. The final score also gives each task equal weight.

This balance does not guarantee useful sharing. A comparison with separate task models, or with one input type removed, would test whether the shared representation transfers knowledge or simply accommodates many outputs. Such controlled comparisons are not yet reported.

Classification rewards combine correct decisions, performance across classes and confidence in the answer. Numerical predictions are judged relative to variation in the training data. The eleven task rewards are then averaged, so improving a small task can matter as much as improving a large one.

Eligibility includes both structured fields and generated text. Its text score measures character overlap, which can miss a reversal of meaning: changing “exclude” to “include” leaves much of a sentence unchanged. A high text score therefore does not establish faithful eligibility criteria.

The full-data reference achieves an average held-out reward of about 0.505. Approval is its strongest task at about 0.615; failure reason is weakest at about 0.309. These are composite rewards, not percentages of clinically correct decisions.

The results show that one trained model can produce all eleven outputs. They do not isolate the benefits of shared learning, chemistry or text. Comparative-effect evaluation has only 49 records, and variation across training runs is unreported, making small differences especially hard to interpret.

Remove one modality at a time and inspect task-level changes. Does chemistry improve safety prediction? Does protocol text help duration? Also test missing inputs and contradictions in generated eligibility criteria. These checks would make the model’s strengths more interpretable than a single average score.