Why Machine Learning Mimics Suboptimal Clinical Decisions

18 concepts13 papers1 proposal

When we train machine learning models on historical Electronic Health Records (EHRs), we often assume we are teaching them the objective language of medicine. We treat clinical decisions—such as ordering a blood test, prescribing an antibiotic, or adjusting a ventilator—as ground-truth labels.

However, a fundamental problem emerges: EHR data does not reflect objective biological truth; it reflects human clinical behavior.

When a model is optimized to minimize prediction error against historical human actions, it inherently learns to replicate the habits, inefficiencies, and errors of the clinicians who generated that data. This phenomenon is known as behavioral mimicry bias. Instead of discovering an optimal path to patient recovery, the model learns to automate the status quo.


The Mechanism of Mimicry

To understand why this happens, consider how standard supervised learning works in clinical settings. If a model is tasked with predicting future clinical events, its loss function penalizes it whenever its prediction deviates from what a human practitioner actually did.

As Choi et al. observed in their early work on temporal clinical event prediction (Doctor AI), models trained on historical data risk learning the average behavior of practitioners rather than identifying optimal medical decisions. If the average practitioner routinely over-orders tests or delays a critical intervention, the model will learn that this delay is the "correct" choice.

This becomes dangerous when human errors are baked directly into the training targets. For example, Shang et al. demonstrated in their work on medication combinations (GAMENet) that simply maximizing prediction accuracy against human-generated ground truth causes models to replicate severe clinical errors, such as prescribing drugs with known harmful interactions. The model cannot distinguish between a brilliant clinical insight and a dangerous oversight; it only knows what is common in the data.


The Objective Truth vs. Behavioral Proxy Dilemma

The root cause of behavioral mimicry bias is a misalignment of training labels. In an ideal world, a model would learn from objective outcomes (e.g., did the patient survive? did the biopsy show malignancy?). In reality, models are frequently trained on behavioral proxies (e.g., what did the doctor prescribe?).

[Patient State] ──> [Clinician Action (Behavioral Proxy)] ──> [Model Training Target]
                          │
                          ▼
             (Includes biases, errors, 
              and systemic inefficiencies)

Cohen et al. explicitly highlighted this distinction, noting that when training labels are based on clinician behavior rather than objective outcomes, models institutionalize subjective biases and suboptimal habits (Problèmes associés au déploiement des modèles fondés sur l’apprentissage machine en santé). This creates a fundamental gap between mimicking current medical practice and optimizing for long-term patient outcomes, a challenge also emphasized by Fouladvand et al. in the context of specialty procedure orders (Graph-Based Clinical Recommender).

When a model's primary objective is to match human behavior, it loses the capacity for innovation. As Liu et al. argued, this focus on mimicry prevents models from discovering truly optimal treatment strategies in high-stress environments like the ICU (Dose Regulation Model of Norepinephrine Based on LSTM Network and Clustering Analysis in Sepsis).


The Danger of Feedback Loops

The problem compounds dramatically when these models are deployed back into active clinical workflows. A mimicking model creates a self-reinforcing feedback loop that can permanently entrench systemic flaws.

┌─────────────────────────────────────────────────────────┐
│                                                         ▼
[Suboptimal Historical Data] ──> [Mimicking Model] ──> [Biased Clinical Recommendation]

In a foundational roadmap for machine learning in health, Schulam et al. explained that integrating predictive models into clinical workflows feeds model-influenced decisions back into future training sets (A Review of Challenges and Opportunities in Machine Learning for Health). This process risks codifying existing clinical biases.

If clinicians begin relying on a model that recommends treatments based on historical averages, they will act on those recommendations, generating new data that validates the model's original bias. Bhattacharyay et al. (TILTomorrow today) and Jeanselme et al. (DeepJoint) both warn that these loops can institutionalize inefficiencies—such as shifting care patterns based purely on weekend staffing schedules—rather than identifying ideal, objective care paths.


Why Reinforcement Learning Doesn't Automatically Fix It

To bypass the limitations of supervised learning, researchers often turn to Reinforcement Learning (RL) to optimize for patient outcomes directly. However, RL agents in healthcare cannot safely explore environments through trial and error; they must learn off-policy from historical data. This reintroduces mimicry bias through the back door.

As Catling et al. noted in a scoping review of cardiovascular resuscitation models, RL agents trained on historical data frequently capture human errors and biases instead of discovering optimal strategies (Data-driven decision support for individualised cardiovascular resuscitation in sepsis). This makes it incredibly difficult to evaluate whether an automated system is genuinely improving care or merely automating standard practice.

Even when models are explicitly designed to balance human guidance with autonomous exploration, the human baseline acts as a powerful gravity well. For instance, Yu et al. developed a supervised-actor-critic framework to safely manage mechanical ventilation (Supervised-actor-critic reinforcement learning for intelligent mechanical ventilation and sedative dosing in intensive care units). They discovered a critical caveat: if the supervisor or reference data reflects inefficient clinical behaviors, the resulting autonomous policy will prioritize mimicking those flaws over finding an optimal path.


Breaking the Mimicry Loop

To move beyond simple behavioral replication, modern clinical AI architectures employ several distinct strategies to decouple clinician habits from patient outcomes:

  • Counterfactual and Synthetic Outcomes: Instead of asking "What did the doctor do?", models can be designed to estimate counterfactual scenarios. Rawson et al. utilized synthetic outcome estimation to model patient trajectories under two hypothetical scenarios: stopping or continuing antibiotics (Machine learning and synthetic outcome estimation for individualised antimicrobial cessation). This allows the model to break free from past prescribing biases by focusing on predicted patient survival rather than historical clinician habits.
  • Expert-Derived Rules and Constraints: Purely data-driven policies cannot distinguish between a necessary clinical intervention and a redundant historical habit. To address routine over-testing in ICUs, Goldenberg et al. combined off-policy learning with expert clinical rules and future-state forecasting (ExOSITO). By forcing the model to adhere to external clinical constraints, they prevented it from replicating historical over-ordering patterns.
  • Interactive Digital Twins: Rather than outputting a single prescriptive recommendation, systems can simulate the trade-offs of various treatment paths. Wentzel et al. developed a visual digital twin system (DITTO) that allows oncologists to explore simulated patient responses to different treatment sequences. This interactive approach uses machine learning to project outcomes while leaving the final decision to the clinician, mitigating the risk of reinforcing outdated historical practices through blind automation.

Key Foundational Work

Doctor AI: Predicting Clinical Events via Recurrent Neural Networks (2015) — contributed 1 concept elements to this direction.
A Review of Challenges and Opportunities in Machine Learning for Health (2018) — contributed 1 concept elements to this direction.
History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 13 papers
  2. Version 1 · 2026-10-04 · 20 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_behavioral_mimicry_bias_d5c91776,
  title = {behavioral mimicry bias},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/behavioral-mimicry-bias/rdr_d5c91776ba55494a9a0a6aaa0c339f86},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 2, frozen text):

@misc{lacuna_behavioral_mimicry_bias_d5c91776_v2,
  title = {behavioral mimicry bias},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/behavioral-mimicry-bias/rdr_d5c91776ba55494a9a0a6aaa0c339f86/v/2},
  note = {Version 2, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake