SCOUT folds 19 side effects into one learned outcome and flags lithium effects in bipolar observational data

In observational biomedical data, treatment effects are often buried under confounding spread across many noisy baseline variables, and it is rarely clear which outcome should capture the response. SCOUT puts both problems into one learning procedure: it builds a low-dimensional adjustment representation, then learns a treatment-responsive composite outcome within it. Earlier practice either prespecified a single endpoint or fixed composite score, or treated the outcome as given while adjusting for confounding; handling adjustment and outcome learning separately lets small effects be lost to multiple-testing penalties and to the power cost of high-dimensional adjustment. In synthetic and semi-synthetic experiments SCOUT's average-effect accuracy was close to CV-TMLE, while CATE error for its score-indexed target was substantially lower; the semi-synthetic data were calibrated to 364 baseline features from the CATIE study.

The method builds adjustment scores once to absorb baseline information, then solves for nonnegative weights on the residualized data, giving one composite outcome shared across follow-up visits. Nonnegative least squares keeps the weights nonnegative because every outcome component is already oriented in a prespecified clinical direction, and negative weights would weaken the composite's clinical interpretability. Once the weights are fixed, two-dimensional adjustment scores are rebuilt at each visit and an ordinary least squares treatment-effect model is fitted; in simulation the algorithm completed in at most 1.2 seconds on average, while CV-TMLE was typically two to three orders of magnitude slower.

Significance is assessed with a conditional randomization test that permutes treatment and reruns the whole procedure, accounting for the uncertainty introduced by adaptively choosing the outcome. The heterogeneity test uses a fourth-order statistic that places more weight on patients whose effects deviate unusually far from the average, rather than looking only at overall variance. Under zero-effect and homogeneous-effect nulls, 95% confidence intervals for empirical Type I error covered the nominal 5% level; in the STEP-BD analysis the heterogeneity test gave a Holm-adjusted value of 0.033.

The demonstrated scope is synthetic and semi-synthetic simulation plus an observational analysis of 481 STEP-BD participants with bipolar disorder (145 lithium initiators and 336 controls); the analysis is restricted to 19 adverse effects observed while patients stayed on their original treatment strategy, at follow-up points 45, 60, 75 and 90 days after the index visit. The authors recommend defining outcome panels around clinically coherent domains rather than indiscriminately including every available outcome.

The authors note that the current method relies on linear representations and a linear treatment-effect model, that the conditional randomization test assumes a fixed treatment assignment across follow-up visits and cannot handle time-varying treatment decisions, and that the initial adjustment representation contains up to as many dimensions per visit as there are outcome components, so its parsimony declines as components grow. The score-indexed CATE retains only part of the full-covariate CATE variance, below 50% in most synthetic settings, and falls as the number of proxies grows.