To train machine learning models for healthcare, researchers require massive datasets containing millions of patient histories. The most accessible repositories of this scale are Electronic Health Records (EHRs) and insurance claims databases. However, these systems were not designed for scientific discovery or clinical precision; they were optimized for administrative tracking and financial reimbursement.
As noted by the authors of A Review of Challenges and Opportunities in Machine Learning for Health and argued by Lawrence Hunter, this creates a fundamental misalignment. Machine learning models treat administrative billing codes (such as ICD-10 codes) as objective biological ground truth, when they are actually the "digital exhaust" of financial transactions and institutional workflows.
The Mechanism: How Billing Distorts Clinical Reality
Administrative codes diverge from a patient's true physiological state through three primary mechanisms: financial optimization, institutional habits, and structural constraints.
1. Financial Optimization (Upcoding and Under-coding)
Medical coding is heavily shaped by insurance reimbursement policies. Hospitals may prioritize, add, or omit certain codes to maximize financial returns or comply with payer guidelines. The authors of HiCu: Leveraging Hierarchy for Curriculum Learning in Automated ICD Coding highlight that datasets inherit a systematic bias toward financial optimization rather than pure clinical description.
This optimization manifests in two ways:
- Upcoding: Adding non-specific or exaggerated secondary diagnoses to justify higher billing tiers, as noted by the authors of Machine learning prediction of incidence of Alzheimer's disease using large-scale administrative health data.
- Under-coding: Omitting conditions that do not impact the immediate reimbursement of a hospital stay. George Hripcsak demonstrates that claims data frequently suffers from severe under-reporting, leading to high false-negative rates for secondary chronic conditions.
2. Institutional and Regional Habits
The assignment of a diagnostic code often depends less on the patient's objective biology and more on where they are treated. The authors of Characterizing the limitations of using diagnosis codes in the context of machine learning for healthcare compared ICD codes directly against objective laboratory data across multiple large cohorts. They discovered that coding practices vary wildly between institutions based on administrative habits and software constraints rather than clinical reality. This institutional heterogeneity is echoed by the authors of Enhancing Prediction Models for One-Year Mortality in Patients with Acute Myocardial Infarction and Post Myocardial Infarction Syndrome and Validation of the Hospital Frailty Risk Score in China, who document regional under-coding and varying record-keeping standards that distort population-level data.
3. Structural and Temporal Noise
In an administrative system, the order and timing of codes reflect billing workflows rather than clinical progression. The authors of Med-BERT: pretrained contextualized embeddings on large-scale structured electronic health records for disease prediction observe that the priority or sequence of medical codes is often driven by financial requirements, scrambling the true clinical hierarchy. Furthermore, the authors of Estimation of Cardiac and Non-cardiac Diagnosis from Electrocardiogram Features note that billing codes lack precise clinical timing, making it difficult to correlate a specific code with real-time physiological measurements.
[True Clinical State] ──> [Financial Incentives] ──┐
──> [Institutional Habits] ├──> [Administrative Code] ──> [Model Training]
──> [Workflow Timelines] ──┘
How Administrative Noise Sabotages Machine Learning
When deep learning models are trained on this noisy data, administrative bias propagates directly into the model's parameters, causing several critical failures.
Learning the System, Not the Disease
Because models lack an inherent mechanism to distinguish between clinical ground truth and administrative processes, they optimize for predicting how a hospital bills rather than how a disease progresses. If a model is trained to predict a condition using billing codes as labels, it may simply learn to detect the financial and administrative patterns that trigger a clerk to enter that code.
Generalizability Collapse
A model that achieves high accuracy in one hospital system often fails when deployed in another. As the authors of Evaluating the performance and potential bias of predictive models for detection of transthyretin cardiac amyloidosis demonstrate, the inconsistency of administrative coding makes claims-based models significantly less reliable and less generalizable across different healthcare networks than models built on direct physiological or imaging data.
Relational Graph Distortion
For advanced architectures like Graph Neural Networks that map disease trajectories, coding inconsistencies degrade the quality of the underlying patient-disease graphs. The authors of A weighted patient network-based framework for predicting chronic diseases using graph neural networks show that variance in coding standards introduces noise into relational graphs, directly degrading downstream predictive accuracy.
Strategies for Grounding Healthcare AI
To build robust clinical models, researchers are developing techniques to bypass or correct administrative coding noise.
Multimodal Grounding
Instead of relying solely on billing codes, modern frameworks integrate raw, objective data streams. For instance, the authors of Enhancing clinical decision support with physiological waveforms - A multimodal benchmark in emergency care incorporate raw physiological waveforms (like ECGs) to anchor diagnostic predictions. Similarly, the authors of Enhancing Prediction Models for One-Year Mortality in Patients with Acute Myocardial Infarction and Post Myocardial Infarction Syndrome leverage unstructured clinical text from discharge summaries to capture the true clinical narrative hidden beneath the billing codes.
Algorithmic Phenotyping
Rather than accepting codes at face value, researchers use clinically-informed time windows and cross-validation with free-text notes or lab results to isolate true phenotypes from administrative noise. This approach, demonstrated by the authors of Curating Retrospective Multimodal and Longitudinal Data for Community Cohorts at Risk for Lung Cancer and Describing the Relationship between Cat Bites and Human Depression Using Data from an Electronic Health Record, ensures that the labels used for model training correspond to verified clinical events rather than administrative proxies.
Iterative, Multidisciplinary Design
Ultimately, mitigating this bias requires moving away from purely algorithmic solutions. As argued by the authors of Challenges and recommendations for high quality research using electronic health records, data scientists must collaborate with clinicians and informaticians to audit the data generation pipeline, ensuring that the financial and software constraints that produced the codes are explicitly accounted for during model training.
Key Foundational Work
Health and Clinical Management (1994) — contributed 3 concept elements to this direction.
Signs of the 2009 Influenza Pandemic in the New York-Presbyterian Hospital Electronic Health Records (2010) — contributed 1 concept elements to this direction.