Beyond Imitation in Medical Machine Learning

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 220 papers29 conceptsfrozen 2026-10-04

When we train a machine learning model on Electronic Health Records (EHRs), we often frame the task as a prediction problem: "Given this patient's history, what will the clinician do next?" While this approach—often called behavioral cloning or imitation learning—can produce models that achieve high accuracy on historical data, it contains a fundamental flaw. It assumes that the historical record represents an "optimal" ground truth.

As researchers like Peter Schulam and Rajesh Ranganath point out in their review of clinical ML challenges, medical data is "messy." It is collected for care, not for analysis, and it is saturated with the subjective biases, errors, and systemic inefficiencies of the humans who generated it.

The Imitation Trap

The core tension in this field is the difference between mimicking a clinician and optimizing for a patient. In early influential work like Doctor AI, Edward Choi and Jimeng Sun demonstrated that recurrent neural networks could predict future clinical events by learning from longitudinal patient trajectories. However, they also noted a critical caveat: these models risk learning the "average" behavior of practitioners rather than identifying the best possible medical decisions.

If a model is optimized solely to replicate past decisions, it becomes a mirror of existing practice—including its flaws. Junyuan Shang and colleagues illustrated this in their work on GAMENet. They found that if a model simply mimics historical prescriptions, it may inadvertently learn to recommend dangerous drug-drug interactions (DDIs) that were present in the training data. To solve this, they had to augment the network with external knowledge graphs to provide a "safety rail" that historical data alone could not provide.

Feedback Loops and Institutionalized Bias

The danger of imitation extends beyond simple errors. When these models are deployed, they create feedback loops. Shubhayu Bhattacharyay and the TILTomorrow team argue that deploying models trained on historical habits can institutionalize those very habits. If a model suggests a treatment because "that's how we've always done it," and clinicians follow that suggestion, the resulting data reinforces the model's original bias.

This is particularly problematic regarding equity. Joseph Paul Cohen and Marzyeh Ghassemi have highlighted that models trained on clinician behavior rather than objective outcomes (like biopsy results or survival) tend to replicate subjective biases. Instead of providing an independent "second opinion," the AI acts as an automated version of the status quo, potentially entrenching disparities in how different patient groups are treated.

From What Was Done to What Should Be Done

To move beyond mere imitation, the research direction is shifting toward models that understand the intent behind decisions or optimize for outcomes directly.

  1. Inverse Decision-Making: Rather than just predicting an action, models like iTransplant, developed by Fergus Imrie and Alihan Hüyük, attempt to learn the underlying "policy" or rationale of clinicians. By making these hidden drivers explicit, researchers can identify where clinical variation stems from expertise versus where it stems from suboptimal shortcuts.
  2. Outcome-Driven Reinforcement Learning: In high-stakes environments like the ICU, researchers are moving toward Reinforcement Learning (RL) to optimize for physiological stability. Finneas J. R. Catling and Aldo A. Faisal note that while RL shows promise for personalizing resuscitation, it must be carefully validated to ensure it isn't just "automating current practice" but actually discovering strategies that improve survival.
  3. Digital Twins and Simulation: Systems like DITTO, proposed by Andrew Wentzel and G. Elisabeta Marai, use "digital twins" to simulate various treatment sequences. This allows clinicians to explore "what-if" scenarios, moving the focus from "what is the most likely next step" to "which step leads to the best long-term trade-off between tumor control and toxicity."

Conclusion

The evolution of clinical machine learning is a transition from descriptive models (what happened?) to prescriptive models (what should happen?). As Jingming Liu and Hui Wang emphasize in their work on sepsis dosage regulation, the goal is to move beyond behavioral cloning. By decoupling a model's objectives from historical human actions and anchoring them in objective patient outcomes, the field aims to create tools that don't just mimic the doctor, but actually assist the patient.

Papers in this version

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 1, frozen text):

@misc{lacuna_beyond_imitation_in_medical_machine_lear_d5c91776_v1,
  title = {Beyond Imitation in Medical Machine Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/beyond-imitation-in-medical-machine-learning/rdr_d5c91776ba55494a9a0a6aaa0c339f86/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}