Framework Targets Overconfident Errors in Clinical Prediction

Clinical prediction models can perform well on average while still making incorrect predictions with high confidence, creating risks when their outputs influence clinical decisions. BMC Medical Informatics and Decision Making has accepted an evaluation of Proximal Error-Based Confidence Adjustment (PECA), a framework designed to reduce confidence around patterns associated with previous model errors. The clinical application uses data from the Alzheimer’s Disease Neuroimaging Initiative and the National Alzheimer’s Coordinating Center. Across simulations and an Alzheimer’s disease trial enrolment example, PECA reduced confident misclassifications while broadly preserving predictive performance, although it also produced fewer high-confidence predictions and requires application-specific choices about how strongly confidence should be adjusted. 

Targeting Confidence Around Previous Errors 

PECA focuses on a model’s confidence rather than relying only on overall measures such as accuracy or area under the receiver operating curve. Clinical classifiers commonly produce probabilities that indicate the estimated likelihood of an outcome. High probabilities may be interpreted as strong evidence and can influence downstream decisions, but models may remain highly confident even when a prediction is wrong. Conventional calibration methods can reduce discrepancies between predicted probabilities and observed risk without necessarily reducing the frequency of high-confidence errors. 

Must Read: Human-AI Co-Design Refines Clinical Prediction Models 

The framework identifies an error space from observations in the training process where predicted classes did not match known outcomes. For a new observation, PECA measures similarity to these historical error regions through a locally adaptive kernel. It then moves the predicted probability towards a neutral confidence level in proportion to that similarity. Predictions that resemble previous errors therefore receive a stronger reduction in confidence, while those further from known error regions are adjusted less. 

The local approach responds specifically to areas where the model has previously failed rather than recalibrating probabilities across all observations. Its smoothing strength adapts to the density of nearby errors. The number of neighbouring errors considered can also be adjusted according to the application, with stricter settings reducing the number of predictions that remain highly confident. PECA is model agnostic and was evaluated with logistic regression, XGBoost and support vector machines. 

Simulations Test Performance Under Data Shift 

Simulation experiments examined how PECA affected prediction reliability when models trained on one dataset were applied to another. The tests included situations with no drift, changes in the distribution of predictors, changes in the relationship between predictors and outcomes and combined shifts. PECA was compared with Kernel Mean Matching, Local Recalibration and a Mahalanobis out-of-distribution adjustment. The simulations used separate training, validation and test data and were repeated 100 times. 

Performance was assessed using accuracy, area under the receiver operating curve, Brier score, conditional misclassification rate and the number of observations receiving high-confidence predictions. Dimensionality was varied to test performance as the number of features increased. Across the different data-shift scenarios, PECA produced the lowest conditional misclassification rate and the best global calibration while maintaining area-under-the-curve performance close to the original model. 

The reduction in confident errors was observed both when training and testing data followed the same process and when concept shift made generalisation more difficult. Sensitivity analyses also found that the relative improvement persisted as the ratio of samples to features changed, although accuracy and discrimination declined across methods in higher-dimensional settings. The improvement came with a trade-off: PECA assigned high confidence to fewer predictions. Other adjustment methods showed a similar tension between retaining high-confidence outputs and reducing errors within that subset. 

Alzheimer’s Trial Example Shows the Trade-Off 

The clinical application examined prediction of future cognitive impairment for potential Alzheimer’s disease trial enrolment. The model used baseline clinical variables including age, sex, education, APOE4 allele count and Montreal Cognitive Assessment scores, with amyloid positivity required for inclusion. The target was progression from normal cognition to mild cognitive impairment or dementia within three years. The application also used different definitions of cognitive status between datasets to mimic the type of data drift that can occur between sites. 

PECA was compared with the unadjusted model and other adjustment approaches. It achieved the lowest conditional misclassification rate in the clinical application, while accuracy remained unchanged compared with the standard model and discrimination was similar. In a hypothetical trial power analysis, the PECA-adjusted model consistently achieved greater statistical power than the standard and unadjusted high-confidence models across the small effect sizes examined. The testing cohort was relatively small, which was identified as an important consideration. 

The approach also has broader constraints. Reducing confidence around historical error patterns means fewer predictions are classified as highly confident and, in the trial setting, may increase screening failures. Performance depends on application-specific parameter choices and assumes that future errors retain some similarity to historical errors. Severe concept drift may reduce effectiveness. PECA is intended for models that already have acceptable average performance and is unlikely to add value when the underlying task produces broadly uncertain predictions. 

PECA offers a way to adjust clinical prediction models around regions where confident mistakes have occurred before, rather than relying on average performance measures alone. In simulations and an Alzheimer’s disease trial enrolment application, the framework reduced high-confidence errors without materially weakening broader predictive performance. Its more conservative behaviour reduces the number of predictions treated as highly confident, creating a deliberate trade-off between confidence coverage and error reduction. The findings support closer attention to local error patterns when assessing prediction reliability, while the method’s parameter sensitivity and dependence on stable error structures remain important constraints. 

Source: BMC Medical Informatics & Decision Making