Can AI handle the missing-data problem in GLP-1 research?

Why a missing measurement can change the result

A medical record does not contain every measurement that a trial protocol might require. Measurements arise through care decisions. As a simple hypothetical example, a clinician might request a biomarker more often when a patient's condition changes. The records with a value can then differ systematically from the records without one.

“Missing not at random”, or MNAR, describes situations in which missingness depends on information that is not fully captured by the observed variables. Filling gaps under an inappropriate assumption can distort a treatment comparison even if the completed values look plausible.

What the AI system actually proposes

Olivia Jullian Parra and colleagues combine two components. CausalFlow-T uses an expert-specified directed acyclic graph and patient history to model longitudinal relationships. A separate evolutionary imputer uses a large language model to propose executable procedures for filling missing data.

The LLM's role is to propose operators that can be evaluated, rather than to supply individual clinical entries as unsupported answers. The pipeline then uses completed inputs for treatment-effect estimation. Its performance depends on both stages.

An explicit causal graph records assumptions about which variables affect others. Writing those assumptions down makes them inspectable; it does not prove they match every relationship in the clinical setting.

Separate the benchmark from the patient application

The authors evaluated synthetic and semi-synthetic benchmarks where the underlying counterfactual outcomes were known. They tested missingness levels from 30% to 80%. The proposed imputer achieved the best pooled rank across the evaluated biomarker and causal metrics.

Known truth makes such benchmarks valuable: researchers can ask whether a method recovers the treatment effect they built into the data. The Swiss healthcare application has no known counterfactual for each patient. Its results therefore require a different level of caution.

A method can reconstruct recorded values well while distorting the relationships needed for causal estimation. The paper examines both kinds of performance, rather than treating accurate value prediction as sufficient evidence of an accurate treatment effect.

What the Swiss GLP-1 comparison estimated

The application included 6,114 adults with type 2 diabetes in Swiss FIRE primary-care records: 2,392 initiating a GLP-1 receptor agonist and 3,722 initiating an SGLT-2 inhibitor.

The pipeline estimated one-year per-protocol weight changes of −4.02 kg and −3.04 kg respectively, yielding a comparative difference of −0.98 kg. Per-protocol here refers to the treatment strategy specified by the analysis. It is a model-based estimate from observed care, not a newly randomised comparison.

The analysis compares drug classes in this cohort. It does not establish that every GLP-1 medicine is preferable to every SGLT-2 inhibitor, and it should not be presented as a general obesity-treatment ranking.

Why the narrow interval is not the whole uncertainty

The reported 95% confidence interval for the comparative difference is −1.01 to −0.96 kg. Its narrowness should not be read as certainty about every part of the analysis. Unmeasured confounding, an incorrect causal graph or a wrong missingness assumption can introduce errors beyond the uncertainty summarised by that interval.

The authors compare the estimate with randomised evidence as a plausibility check. Agreement is useful context, but it cannot independently establish the truth of the unobserved outcomes reconstructed from these records.

This is related to the uncertainty problem in GLP-1 digital-twin predictions: a numerical answer may be precise within a model while still depending materially on its assumptions.

What readers should look for in the next validation

Further evaluation should ask whether the method preserves treatment-effect estimates across other patient populations, measurement patterns and plausible causal graphs. Prospective or external comparisons would help show whether the benchmark gains carry into new settings.

The useful contribution is a testable workflow for incomplete longitudinal data. Its practical value will depend on transparent sensitivity analyses and on demonstrating when the reconstructed data support a reliable comparison. None of these findings changes the research-use status of products sold by Remy Research.