Look beyond a single prediction.

How this reconstruction study works

The question: if we hide a value that is already known, can a trained model estimate it from the information left visible? This page reports a saved artificial-data experiment, not a new analysis of patients.

How the model estimates a missing value

The method is a conditional variational autoencoder (CVAE): a neural network trained beforehand to represent relationships among clinical fields and reconstruct their values. “Conditional” means that known profile information helps guide the estimate. It processes the modelled clinical fields, not CT images or the subject identifier.

1. Hide the answer, keep the context. The experiment replaces the selected value with a missing marker. For example, when PD-L1 is hidden, available age, smoking, stage and other modelled fields remain as context. The original PD-L1 is kept separately for scoring; it is not supplied as an input.
2. Tell the network what is known. Numerical values are scaled using saved training statistics and categories are encoded. A separate observed/missing mask tells the network which entries carry information. Internal zero placeholders for missing values do not mean an actual age or PD-L1 of zero, and are not final mean-imputation answers.
3. Summarise the profile and decode an estimate. The encoder combines the visible fields and conditions into a compact learned representation. The decoder uses that representation to predict missing values. At reconstruction time the implemented pathway uses the encoder’s mean representation, rather than drawing a new random profile as cohort generation does.
4. Return a value or a category. Numerical estimates are returned in the field’s original units, with the configured bounds and rounding. For categorical fields, the highest-scoring category is selected. That score is a raw model probability, not calibrated confidence. Observed values are copied unchanged; missing modelled fields are filled. Only deliberately hidden values with known originals contribute to these accuracy/error measurements.

This is a learned estimate, not recovery of a hidden fact, a lookup of the nearest patient, or a single median assigned to everyone. Different visible profiles can yield different estimates. Limited information can still favour central numerical values or frequent categories, which is why class recall and errors across value ranges matter.

Which experiments are shown?

The diagnostic uses the first 50 subjects from the service’s ordered artificial evaluation set, not 50 real patients. Eight fields were tested one at a time. Four additional tasks hid age + PD-L1; smoking + ECOG; the lung + brain metastasis flags; or age + PD-L1 + smoking + ECOG. A subject contributes to a task only when every deliberately hidden field has a known original value, so denominators differ.

The routine loads an existing model and calls its reconstruction method; it does not train it during this evaluation. It scores predictions directly, without the separate consistency-repair stage used in cohort generation. Opening this page reads the stored aggregate report and runs neither training nor reconstruction.

Why reconstructing several fields deserves special attention

When several fields are hidden together, the model has less observed information to work with and must estimate several related values. This is generally a more demanding problem than reconstructing one missing field. It does not mean every multiple-mask score must be worse: the selected variables, eligible cases and learned relationships can produce different outcomes.

Future development priority: strengthen reconstruction when several clinical fields are missing, and evaluate accuracy and consistency across different missingness patterns, subgroups and authorised datasets. The comparisons below identify where to focus that work; they do not yet demonstrate that this broader problem has been solved.

What did the stored study show?

What these results can establish

They demonstrate and measure a reconstruction workflow on artificial data. They do not establish performance on real clinical records, clinical validity, privacy protection or an official KPI result. Hiding known values is a controlled test; real-world missingness may behave differently. Multiple-mask versus single-mask differences are descriptive, not a general causal claim.

Try a single artificial example in Reconstruct → That interactive request does not update this stored study. Continue below for all field results, eligible counts, class matrices and numerical ranges.