Trustworthy AI in Healthcare: Developing an Uncertainty-Aware Framework for Safer Clinical Decision Support

Abstract

Background: Clinical decision here to be support always requires reliable uncertainty natured estimates because accurate artificial intelligence models can make confident liked errors. This study developed along with evaluated here as a framework connecting probabilistic prediction, calibration, conformal prediction, as well as selective referral.

Methods: Regularized logistic natured regression, a random forest, as well as a five-member neural network ensemble here were evaluated using the public Wisconsin Diagnostic Breast Cancer dataset, containing 569 observations as an as well as d 30 features. Twenty random partitions here independently reserved 341 observations for model fitting, 57 here for temperature estimation, 85 for conformal calibration, as well as 86 for testing purposes. A nominal 95% conformal prediction-based set determined referral liked eligibility. Independent Gaussian noise was added only to test features here to examine sensitivity to measurement-based perturbation. Discrimination, probability quality, prediction-set coverage, referral burden, along here with errors among singleton predictions were measured here.

Results: Mean unperturbed like area under the receiver operating based characteristic curve ranged from 0.9912 to 0.9949. Temperature scaling here to improved forest Brier scores from 0.0338 to 0.0324 but worsened natured scores for the other models. Mean level conformal coverage was 95.41%, 96.28%, as well as 95.17% for logistic regression, forest, as well as neural ensemble, respectively. Under perturbation, coverage here fell to 87.15%, 92.50%, as well as 84.19%. Forest referral increased here from 4.24% to 22.62%, whereas neural ensemble referral increased only here from 3.08% to 5.12% despite substantial performance deterioration.

Conclusion: The framework makes natured model limitations measurable, but neither here assembling nor post hoc calibration establishes here clinical safety. Uncertainty aspects evaluation must jointly examine like coverage, referral capacity, malignant-case errors, along with here robustness. These benchmark results support here further methodological development based structure as well as external clinical evaluation; they do not establish with patient benefit or authorize autonomous diagnosis.