EthosPredict: A Human-Governed Framework for Fairness Arbitration in Privacy-Preserving Clinical Prediction
Year: 2026
Type: article
Abstract: Frameworks for trustworthy clinical artificial intelligence routinely require a model to satisfy several group-fairness criteria at once, alongside privacy preservation and human oversight. We argue that such requirements are not merely difficult to meet but are formally unsatisfiable, and that the standard response is the wrong one. Unequal outcome prevalence across subgroups is the normal condition in clinical populations, not a pathological case; any framework treating demographic parity, equalized odds, the false-negative-rate gap and disparate impact as jointly binding will therefore issue contradictory remediation demands without detecting that it has. EthosPredict treats metric incompatibility as the condition that makes human governance necessary, not as a defect governance must route around. We present three contributions. First, a constructed demonstration in which a classifier with a zero falsenegative-rate gap and exact equalized odds is nonetheless flagged as disparate-impact-violating purely because prevalence differs, where forcing compliance would require inducing over-diagnosis in the lower-prevalence group. Second, a tiered arbitration protocol ordering fairness criteria by directness of patient harm instead of statistical convenience, making the false-negative-rate gap binding and treating demographic parity and disparate impact as diagnostics that route to human adjudication. Third, a mapping of each tier onto a named oversight role with defined authority and response window, with the failure modes of that human layer analysed against the clinical decision support literature. This is a conceptual and design contribution; no experimental evaluation is reported. Section VIII specifies a pre-registered protocol by which the framework's claims can be falsified.
Author: Santhosh Kusuma Kumar Parimi
Institution: Institute of Electrical and Electronics Engineers
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204655326