Deceptive bias measurement in deep learning: Assessing shortcut reliance in TCGA cancer models
- 1Department of Electrical, Computer and Software Engineering, Ontario Tech University, Oshawa, Ontario, Canada
- 2Department of Engineering, Brock University, St. Catharines, Ontario, Canada
Abstract
Machine learning bias is a persistent challenge because it can create unfair outcomes, limit generalization, and reduce trust in real-world applications. A key source of this problem is shortcut learning, where models exploit signals linked to sensitive attributes, such as the data source or collection site, instead of relying on task-relevant features. To address this, we propose the Deceptive Signal metric, a novel quantitative measure designed to assess the extent of a model’s reliance on hidden shortcuts during the learning process. This metric is derived via the Deceptive Bias Detection pipeline, which isolates shortcut dependence by contrasting the model’s behavior under two controlled conditions: (1) Full Exclusion, where a sensitive subgroup is completely removed from training; and (2) Partial Exclusion, where the model has limited access to specific classes within that subgroup. By calculating the behavioral shift between these settings, the Deceptive Signal metric provides a quantitative estimate of the model’s susceptibility to learning task-irrelevant patterns. In experiments with the TCGA histopathology dataset, our metric successfully quantified substantial dependencies on center-specific artifacts in models trained for cancer classification. Author summary: Deep learning models are becoming powerful tools in healthcare, but they often suffer from a critical vulnerability: they can get the right answer for the wrong reason. In medical imaging, an AI might correctly identify a tumor not by analyzing the tissue, but by recognizing irrelevant digital markers unique to the specific hospital or scanner that produced the image. This phenomenon, known as shortcut learning, makes AI systems appear accurate at first glance while potentially limiting reliability in real-world patient care settings.To solve this, our research moves beyond simple accuracy checks and introduces a specific quantitative metric for shortcut learning. We developed a testing framework that forces the model into controlled training scenarios, deliberately withholding specific shortcut information to see how the model reacts. By mathematically comparing the model’s behavior across these scenarios, we calculate a quantitative score that indicates the magnitude of the model’s dependence on irrelevant patterns. This metric allows us to put a concrete number on the model’s trustworthiness, helping assess whether predictions are more strongly driven by task-relevant biological features rather than background artifacts.
Related
- Unstable word form representations and greater reliance on right hemisphere during word learning in adults with a history of developmental language disorder
- High-throughput mitochondrial aerobic metabolism measurement: assessing a plate-reader approach using avian blood cells
- Assessing measurement uncertainty at laboratory network scale and its impact on diagnostic performance in the absence of a gold standard: application to ELISA tests for Coxiella burnetii in ruminants
- Floral scents of deceptive Arum maculatum L. (Araceae) shape a taxonomically diverse visitor community of lesser dung flies (Sphaeroceridae) that affects its reproductive success
- Loss contexts enhance dorsolateral prefrontal interpersonal neural synchrony during successful human deceptive recommendations
BibTeX
@article{kheiri2026deceptive,
title = {{Deceptive bias measurement in deep learning: Assessing shortcut reliance in TCGA cancer models}},
author = {Farnaz Kheiri and Shahryar Rahnamayan and Masoud Makrehchi},
journal = {PLOS Digital Health},
year = {2026},
month = sep,
doi = {10.1371/journal.pdig.0001165},
url = {https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0001165},
note = {CC BY 4.0},
}
RIS
TY - JOUR
TI - Deceptive bias measurement in deep learning: Assessing shortcut reliance in TCGA cancer models
AU - Farnaz Kheiri
AU - Shahryar Rahnamayan
AU - Masoud Makrehchi
PY - 2026
DA - 2026/09/03
JO - PLOS Digital Health
DO - 10.1371/journal.pdig.0001165
UR - https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0001165
AB - Machine learning bias is a persistent challenge because it can create unfair outcomes, limit generalization, and reduce trust in real-world applications. A key source of this problem is shortcut learning, where models exploit signals linked to sensitive attributes, such as the data source or collection site, instead of relying on task-relevant features. To address this, we propose the Deceptive Signal metric, a novel quantitative measure designed to assess the extent of a model’s reliance on hidden shortcuts during the learning process. This metric is derived via the Deceptive Bias Detection pipeline, which isolates shortcut dependence by contrasting the model’s behavior under two controlled conditions: (1) Full Exclusion, where a sensitive subgroup is completely removed from training; and (2) Partial Exclusion, where the model has limited access to specific classes within that subgroup. By calculating the behavioral shift between these settings, the Deceptive Signal metric provides a quantitative estimate of the model’s susceptibility to learning task-irrelevant patterns. In experiments with the TCGA histopathology dataset, our metric successfully quantified substantial dependencies on center-specific artifacts in models trained for cancer classification. Author summary: Deep learning models are becoming powerful tools in healthcare, but they often suffer from a critical vulnerability: they can get the right answer for the wrong reason. In medical imaging, an AI might correctly identify a tumor not by analyzing the tissue, but by recognizing irrelevant digital markers unique to the specific hospital or scanner that produced the image. This phenomenon, known as shortcut learning, makes AI systems appear accurate at first glance while potentially limiting reliability in real-world patient care settings.To solve this, our research moves beyond simple accuracy checks and introduces a specific quantitative metric for shortcut learning. We developed a testing framework that forces the model into controlled training scenarios, deliberately withholding specific shortcut information to see how the model reacts. By mathematically comparing the model’s behavior across these scenarios, we calculate a quantitative score that indicates the magnitude of the model’s dependence on irrelevant patterns. This metric allows us to put a concrete number on the model’s trustworthiness, helping assess whether predictions are more strongly driven by task-relevant biological features rather than background artifacts.
SN - 2767-3170
ER -