Learning from Incomplete Electronic Health Records

51 concepts44 papers2 proposals

Machine learning models thrive on complete, well-structured matrices. In an ideal dataset, every patient has a complete set of vitals, a standard battery of lab results, and a clearly documented clinical trajectory.

Real-world Electronic Health Records (EHRs), however, look more like Swiss cheese. Gaps in data are the norm rather than the exception. As Surati et al. note in their work on translational databases, manual retrospective data entry is an incredibly labor-intensive process that naturally introduces missing fields. Furthermore, Rudrapatna et al. point out a fundamental tension in medical informatics: clinicians prioritize immediate patient care over documenting systematic metrics for future research.

When machine learning models encounter these real-world data gaps, standard practices fail. Simply discarding incomplete patient records creates an artificial training environment. As Rosnati et al. demonstrate, models trained on perfectly curated benchmarks frequently collapse when deployed in actual clinical settings where decisions must be made despite missing labs. To build clinical models that work in the real world, we have to understand why data goes missing and design algorithms that treat missingness not as a failure of data collection, but as an informative feature.

The Documentation Paradox and Informative Missingness

In standard statistical applications, missing data is often treated as "missing at random"—a random glitch in data collection. In healthcare, missingness is rarely random; it is driven by clinical workflows, resource constraints, and patient outcomes. This phenomenon is known as informative missingness.

Consider a striking example discovered by Hripcsak et al. termed the documentation paradox. When mining EHR data, naive automated approaches can produce highly counterintuitive results, such as showing that the "healthiest-appearing" patients have the highest mortality rates. Why does this happen?

[ Patient Admits with Severe Trauma ] 
               │
               ▼
[ Rapid Mortality / Immediate Surgery ] ──► Minimal Time for Documentation ──► "Empty" EHR Profile ──► Naive Model Classifies as "Low Risk"

When a patient is critically ill or experiences rapid mortality, clinicians focus entirely on life-saving interventions rather than filling out electronic forms. Because the patient dies or enters surgery quickly, very few records are generated. A naive algorithm looks at this sparse profile, interprets the lack of documented symptoms as an absence of illness, and misclassifies a high-risk patient as low-risk.

Informative missingness also manifests through routine diagnostic choices. As Williams et al. observe when calculating clinical severity scores retrospectively, clinicians only order laboratory tests relevant to a patient's immediate presentation. If a specific lab value is missing, it is often because a physician judged that the patient did not need it. Assuming this data is missing at random leads to a systematic underestimation of organ dysfunction and disease severity across retrospective cohorts.

Similar patterns emerge across various clinical contexts:

  • Treatment Discontinuation: In observational studies, records often omit why a patient stopped a treatment, making it difficult to differentiate between treatment failure and a patient's personal choice (Nieva et al.).
  • Bedside Procedures: Point-of-care ultrasounds performed rapidly at the bedside are frequently left unlogged in the central EHR, obscuring true imaging utilization trends (Chang et al.).
  • Longitudinal Noise: Missing tumor stages or molecular markers in free-text pathology reports introduce significant noise into survival analyses (Kurant et al.).

Beyond Naive Imputation: Algorithmic Solutions

When developing predictive models, researchers face a dilemma: how do you handle these gaps without introducing bias? Traditional approaches like mean imputation distort the relationships between variables, undermining model reliability (Maffezzoni et al.; Shi et al.). Modern techniques solve this by either imputing data intelligently using global context or bypassing imputation entirely.

Joint Imputation and Classification

Rather than treating imputation as a disconnected preprocessing step, modern architectures integrate it directly into the predictive task. For instance, Vivar et al. developed Multigraph Geometric Matrix Completion (MGMC). Instead of guessing missing values in isolation, MGMC leverages population-based graphs to simultaneously fill in missing clinical metrics and predict disease classification, ensuring that the imputed values are mathematically optimized to preserve diagnostic relationships.

For sparse, non-ICU temporal data, Dai et al. introduced a similarity-aware diffusion model (SADI). Recognizing that a patient's timeline is often interrupted by irregular visits, SADI uses a 3D-attention mechanism to "borrow" missing temporal data from cohorts of patients with highly similar clinical profiles, outperforming traditional autoregressive models.

Dynamic, Reduced-Feature Modeling

An alternative strategy is to avoid imputation altogether and build models that adapt on-the-fly to whatever data happens to be available.

In scenarios of "extreme missingness"—where over half of the expected clinical features are absent—McCombe et al. demonstrated that a "reduced-feature" classification approach is highly effective. Instead of trying to guess the missing values, the system dynamically reconstructs the machine learning model using only the features present at that moment, balancing diagnostic speed and accuracy.

Available Features: [ Age, Blood Pressure, Glucose ]  ──►  [ Dynamic Model Generator ] ──► Custom Architecture Optimized for 3 Features
Missing Features:   [ BMI, Cholesterol ]

Taking this concept a step further, Ardulov et al. reformulated clinical diagnosis as a sequential decision-making process using Q-learning. Instead of assuming all patient data is available simultaneously, their reinforcement learning agent evaluates the patient step-by-step. It actively decides whether to "ask" for a missing feature or make a diagnostic prediction based on the current incomplete state, penalizing excessive or costly data collection.

Recovering Hidden Data from Unstructured Text

Often, data that appears "missing" from structured EHR tables isn't missing from the hospital entirely—it is simply trapped in unstructured text, such as clinical narratives, radiology reports, or progress notes.

For example, Yang et al. highlight that critical patient risk factors, like quantitative smoking status, are routinely omitted from standardized EHR fields. To recover this data, the authors developed specialized Natural Language Processing (NLP) tools to parse informal medical text, effectively transforming narrative descriptions into structured predictors.

Similarly, in medical imaging, quantitative metrics like tumor volume or diameter are frequently left out of structured databases, leading to the exclusion of valuable training samples. To circumvent this, Bassi et al. introduced a training paradigm called Report-Supervision (R-Super). Instead of relying on scarce, manually drawn tumor masks or missing quantitative text attributes, R-Super converts unstructured radiology text directly into voxel-wise loss functions to supervise tumor segmentation in CT scans.

Finally, multimodal models help ground ambiguous or incomplete data points. As Shaaban et al. point out, structured EHRs often contain lab values without accompanying reference thresholds, making it impossible for a model to know if a number is alarming. By combining these unstructured values with visual grounding from chest X-rays, multimodal frameworks can infer the clinical context that the structured data lacks.

Conclusion

Data incompleteness is an inescapable property of real-world medicine. Gaps in patient records are not random errors; they are reflections of clinical urgency, diagnostic reasoning, and documentation workflows. By shifting away from naive data-dropping and moving toward architectures that recognize informative missingness, leverage unstructured text, and dynamically adapt to available features, researchers can build robust clinical models capable of operating safely in imperfect environments.

Key Foundational Work

Using intermediate states to improve the ability of the Arden Syntax to implement care plans and reuse knowledge. (1995) — contributed 1 concept elements to this direction.
Proteomic characterization of non-small cell lung cancer in a comprehensive translational thoracic oncology database (2011) — contributed 1 concept elements to this direction.
Bias Associated with Mining Electronic Health Records (2011) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_retrospective_data_incompleteness_0825f2c5,
  title = {retrospective data incompleteness},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/retrospective-data-incompleteness/rdr_0825f2c51e5d4a5b87f6bdfafa62680d},
  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_retrospective_data_incompleteness_0825f2c5_v1,
  title = {retrospective data incompleteness},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/retrospective-data-incompleteness/rdr_0825f2c51e5d4a5b87f6bdfafa62680d/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake