How Much Representational Work Does a Neural Network Require for a Clinical Prediction?
Proceedings of the AAAI Symposium Series · 06.11.2026 · 10 (1) · 1186-1194
DOI: 10.1609/aaaiss.v10i1.43543 ↗Kurz gefasst
Fachartikel in der Zeitschrift Proceedings of the AAAI Symposium Series, online erschienen am 06.11.2026. Verfasst von Muhammad Umair Danish und 2 weiteren Autor:innen.
Abstract (Originalsprache)
Clinical neural classifiers provide predictions, while existing explainable artificial intelligence (XAI) and uncertainty methods help identify influential inputs and characterize model confidence. However, they provide limited insight into how much internal representational effort a neural network requires to reach an individual decision. We introduce Neural Decision Work (NDW), a post-training method that measures this internal effort during decision formation. NDW maps successive hidden representations into a common diagnostic evidence space defined by training-derived class prototypes and measures the normalized trajectory traversed as the prediction evolves across network depth. Complementary measures characterize the direction, depth, efficiency, and task dependence of this internal work. We evaluate NDW on matched binary and six-class chronic disease classification tasks comprising 406,923 respondents from the 2024 Behavioral Risk Factor Surveillance System (BRFSS). The analysis reveals two key findings. First, multiclass task framing reduces mean NDW by 48.7% relative to binary framing (Wilcoxon rank-biserial correlation =-0.978; n=406,925 paired patient–seed evaluations), indicating that the internal work required for the same patients can vary substantially with the formulation of the prediction task. Second, 8.5% of held-out evaluations fall simultaneously within the top decile of both prediction confidence and NDW, indicating that highly confident outputs can still emerge from comparatively demanding internal representational trajectories. These findings establish NDW as a complementary process-level audit method that captures information not conveyed by output confidence alone.
Zitieren
APA
Danish, M. U., Rehman, U., & Narayan, A. (2026). How Much Representational Work Does a Neural Network Require for a Clinical Prediction?. Proceedings of the AAAI Symposium Series, 10(1), 1186-1194. https://doi.org/10.1609/aaaiss.v10i1.43543
Vancouver
Danish MU, Rehman U, Narayan A. How Much Representational Work Does a Neural Network Require for a Clinical Prediction?. Proceedings of the AAAI Symposium Series. 2026;10(1):1186-1194. doi:10.1609/aaaiss.v10i1.43543
BibTeX
@article{Danish2026,
author = {Danish, Muhammad Umair and Rehman, Umair and Narayan, Apurva},
title = {How Much Representational Work Does a Neural Network Require for a Clinical Prediction?},
journal = {Proceedings of the AAAI Symposium Series},
year = {2026},
volume = {10},
number = {1},
pages = {1186--1194},
doi = {10.1609/aaaiss.v10i1.43543}
}
EUVIMED