Training clinical artificial intelligence models exclusively on the practice patterns of top-performing physicians could eliminate decision noise while preventing algorithms from encoding systemic biases, according to a perspective published in Intensive Care Medicine by Daniel Ostrovsky, Daniel Talmor, and Victor Novack. The authors propose that rather than learning from all cases across an entire clinical staff, machine learning algorithms designed for critical care environments should selectively mimic locally identified best-practice clinicians to improve the quality of automated recommendations.
Medicine operates as a noisy decision-making environment, with noise defined by Kahneman, Sibony, and Sunstein as undesirable variability in judgments of the same problem. This unwarranted clinical variation arises frequently in discretionary decisions where medical evidence fails to point to a single correct action, leaving choices to depend primarily on individual clinician judgment, local practice styles, and patient preferences.
Critical care medicine exhibits widespread discretionary variation across multiple interventions, including hospital intensive care unit admissions that vary substantially even after adjusting for patient case mix and measured hospital factors. Parallel discrepancies appear in routine antibiotic prescribing, the timing and threshold established for invasive mechanical ventilation in hypoxaemic respiratory failure, and the deployment of costly organ-support therapies such as renal replacement therapy and extracorporeal membrane oxygenation.
Current artificial intelligence tools trained on routine outcome data can eliminate decision noise while simultaneously reproducing clinical bias. A uniformly applied algorithm successfully eradicates noise at the level of its recommendations, but whether it removes bias depends entirely on the underlying training data. In current clinical practice, algorithms either learn to reproduce diagnostic labels in settings such as digital pathology and image-based diagnosis, or they train on clinical decisions and patient outcomes when definitive diagnostic labels do not exist.
When algorithms train on operational endpoints such as intensive care unit transfers or clinical deterioration, they rely on data points that are not purely biological. These records inherently reflect clinician practice patterns, local system constraints, and potentially suboptimal care decisions. By mathematically regressing outcomes over an averaged clinical approach, standard models merely reproduce the central tendency of observed practice rather than an optimal standard of care.
The operational hazards of training on unrefined clinical decisions appear clearly in intensive care unit triage, where predictive models are trained on historical admission records rather than identifying which patients would truly deteriorate without specialized care. Because individual physicians exhibit divergent admission thresholds, training a model on unselected institutional datasets causes recommendations to converge on the unit mean, standardizing care around an unvetted average.
In a hypothetical intensive care unit where clinician admission rates range from 0 percent to 24 percent with a mean of 12 percent, an algorithm trained across all clinicians eliminates noise by driving all practitioners toward that 12 percent rate. However, if best-practice clinicians maintain an 8 percent admission rate with non-inferior patient outcomes, standardizing around the institutional mean perpetuates a four-percentage-point over-admission bias, converting sound discretionary judgments into unnecessary and costly admissions.
To overcome this limitation within structured artificial intelligence development phases, Ostrovsky and colleagues suggest filtering data during initial training selection so algorithms learn only from best-practice clinicians. These clinicians could be identified a priori using objective clinical metrics such as low complication rates, cost-benefit metrics like reduced intensive care utilization with non-inferior outcomes, or comprehensive measures including peer rankings, patient assessments, experience, and knowledge tests.
Implementing a selective training framework presents critical methodological hurdles, beginning with the risk that seemingly superior performance stems from non-random case mix rather than better decision-making. Developers must apply rigorous retrospective risk-adjusted comparison methods evaluated against comprehensive outcome matrices, including positive controls, negative controls, and operational performance metrics.
Because intensive care requirements depend heavily on ward capabilities, staffing levels, geographic setting, and physical unit capacity, models trained on top clinicians cannot serve as universal standards without local training or calibration. Furthermore, clinical workflows evolve over time in response to the algorithms themselves, making continuous model validation and drift monitoring mandatory throughout deployment.
Restricting training datasets to a subset of top clinicians reduces overall sample size and narrows data distributions, risking degraded model accuracy in rare patient subgroups. Mitigating these gaps requires federated data access across multiple intensive care units, formal fairness audits, prospective evaluations of clinician behavior and patient outcomes, and robust institutional safeguards against automation bias.
The conceptual separation between decision noise and systematic bias represents the foundation of the authors’ critique of current machine learning methodologies in critical care. In their analysis, Kahneman, Sibony, and Sunstein define noise as unwarranted variability in judgments concerning the identical clinical problem. When an artificial intelligence algorithm is deployed uniformly across an intensive care service, it inherently eliminates noise at the recommendation level because identical patient feature inputs generate consistent algorithmic outputs. However, eliminating noise does not guarantee the removal of bias, as the algorithm’s guidance remains tethered to the underlying data on which it was trained. In domains such as digital pathology and image-based diagnosis, machine learning models are trained to reproduce definitive diagnostic labels, allowing them to simultaneously reduce both noise and diagnostic bias. Conversely, when objective ground-truth labels are absent, algorithms are routinely trained on recorded clinical decisions, such as intensive care transfers, or on clinical endpoints, such as patient deterioration or death, that directly reflect preceding clinical choices. Because these operational outcomes are not purely biological events but instead embody individual practitioner habits and institutional care structures, standard algorithms merely compute mathematical regressions over the prevailing average clinical practice.
This reliance on historical clinical practice patterns introduces acute vulnerabilities in discretionary critical care decisions, where clinical evidence does not define a single correct therapeutic action. In such discretionary settings, clinical choices vary extensively based on local practice styles, individual judgment, and patient preferences, resulting in substantial unwarranted variation across hospital intensive care unit admissions that persists even after rigorous adjustment for patient case mix and measured hospital characteristics. Identical patterns of unwarranted variation characterize routine antibiotic prescribing, the timing and threshold established for invasive mechanical ventilation in hypoxaemic respiratory failure, and the utilization of resource-intensive organ-support modalities like renal replacement therapy and extracorporeal membrane oxygenation. Furthermore, this variability manifests not only between different clinicians but within individual practitioners over time, as a single physician may choose to admit a patient on one day yet refuse admission to an identical patient on another day. Consequently, training triage models on historical admission records captures and reproduces these inconsistent subjective judgments rather than identifying the biological subset of patients who would genuinely deteriorate in the absence of intensive care support.
To demonstrate the mathematical consequences of standardizing care around average clinical practice, Ostrovsky, Talmor, and Novack evaluate the distribution of physician-level admission rates within a hypothetical intensive care unit. In this modeled unit, individual clinician admission rates span from zero percent to 24 percent, forming an observational distribution centered around a unit mean of 12 percent. If an artificial intelligence recommendation algorithm is developed using the complete institutional dataset and clinicians adhere perfectly to its guidance, every practitioner’s admission frequency converges directly upon the 12 percent average, successfully eliminating discretionary noise. However, if the institutional best-practice clinicians, specifically those achieving non-inferior clinical outcomes among both admitted and non-admitted cohorts, maintain an admission rate of only 8 percent, algorithmic convergence upon the unit mean preserves a four-percentage-point over-admission bias. Under these operating conditions, algorithmic standardization actively converts sound clinical judgments into suboptimal decisions by nudging clinicians whose lower admission thresholds are clinically appropriate into admitting patients who could otherwise be safely and effectively managed in standard hospital wards.
Addressing these structural distortions requires intervention at the data selection phase within the structured artificial intelligence development framework established by Van de Sande and colleagues. Following the initial stages of defining the clinical problem and confirming organizational need, developers must deliberately restrict training datasets to cases managed by best-practice clinicians identified a priori according to the target clinical challenge and stakeholder interests. This prior identification can leverage clinical outcome metrics, such as achieving the lowest patient complication rates, or cost-benefit criteria, such as reduced intensive care utilization coupled with non-inferior patient outcomes. Alternatively, selection can incorporate peer rankings, patient assessments, objective knowledge-test performance, clinical experience, or a multidimensional composite of these evaluative measures. In critical care triage, training algorithms exclusively on practitioners who achieve lower admission rates without compromising outcomes across both admitted and non-admitted patient populations enables machine learning tools to eliminate unwarranted noise while simultaneously eradicating over-admission bias.
Implementing this selective strategy also introduces significant statistical and governance hurdles that require rigorous institutional safeguards. Restricting training data to a subset of top clinicians reduces sample size and narrows the training distribution, impairing model accuracy in patient subgroups those clinicians rarely encounter. Mitigating these gaps necessitates broader, more diverse datasets through initiatives like federated data access across intensive care units, alongside formal fairness and subgroup audits. Additionally, deploying models as novel clinical interventions demands prospective evaluation of workflow, clinician behavior, and patient outcomes rather than relying solely on predictive metrics. Health systems must also establish clinical artificial intelligence literacy, defined oversight frameworks, and defenses against automation bias. Ultimately, future empirical investigations must determine whether best-practice clinicians can be reliably identified and whether algorithms trained on their practice patterns yield superior patient outcomes compared to models trained across entire clinical staffs.
Subject of Research: Training clinical artificial intelligence models on top-performing physicians to reduce noise and bias in intensive care
Article Title: AI – Train with the best
Article References: Ostrovsky, D., Talmor, D., & Novack, V. (2026). AI – Train with the best. Intensive Care Medicine. https://doi.org/10.1007/s00134-026-08660-7
Image Credits: AI Generated
DOI: 10.1007/s00134-026-08660-7
Keywords: clinical artificial intelligence, intensive care medicine, unwarranted clinical variation, clinical decision noise, algorithmic bias, ICU triage, intensive care unit admission, best-practice clinicians, invasive ventilation, extracorporeal membrane oxygenation, renal replacement therapy, risk-adjusted comparison, model drift monitoring, federated data access, automation bias, subgroup fairness audits, prospective clinical evaluation
News Source: Blake Davidson. (November 12, 2026). Training Artificial Intelligence on Top Physicians May Reduce Intensive Care Bias. Scienmag.
