📊 statistics

Predicting Hospitalization from a Whole-Person Health Score with Incomplete Electronic Health Records Data: A Case Study

This study demonstrates that predicting hospitalization from incomplete Electronic Health Records using an Allostatic Load Index is most effective when tailoring models to specific patient missing-data patterns, achieving an AUC of 0.73 in-sample, though cross-validation performance drops to 0.63.

Original authors: Grayson E. Weavil, Joseph Rigdon, Sarah C. Lotspeich

Published 2026-06-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Grayson E. Weavil, Joseph Rigdon, Sarah C. Lotspeich

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: A "Health Report Card" with Missing Grades

Imagine you want to give a student a single "Health Report Card" score based on ten different subjects (like Math, History, Science, etc.). This score, called the Allostatic Load Index (ALI), tells you how much stress the student's body is under.

The Goal: If the score is high (many "failing grades"), the student is at risk of getting sick or needing to go to the hospital.
The Problem: In a real classroom, you don't always have test scores for every subject. Maybe the student skipped the Science test, or the History teacher lost the paper. In the real world (Electronic Health Records), doctors don't order every single blood test for every patient. Some data is missing.

The researchers asked: How do we calculate this "Health Report Card" accurately when some of the grades are missing, and can we use it to predict who will end up in the hospital?

The Experiment: 1,000 Students and a Missing Puzzle

The team looked at data from 1,000 patients at a hospital in North Carolina. They had ten specific health markers (like blood pressure, cholesterol, and blood sugar) that make up the ALI score.

The Reality: Some markers were easy to find (like blood pressure, which is taken at almost every visit). Others were very hard to find (like specific blood tests that are only ordered if a doctor suspects a problem).
The Challenge: If you just ignore the missing tests, you might accidentally think a patient is healthy when they aren't. If you guess the missing tests are "bad," you might think they are sicker than they are.

The Strategies: How to Handle the Missing Grades

The researchers tried different ways to fill in the blanks or work around them, testing them like different strategies in a game:

The "Proportion" Method (The Partial Report Card):
Instead of counting missing tests as "good" or "bad," they just calculated the percentage of available tests that were unhealthy.
- Analogy: If a student took 5 out of 10 tests and failed 2, their score is 40% failure, not 20% (which you'd get if you counted the 5 missing tests as passing).
The "Best/Worst Case" Guessing:
- Best Case: Assume all missing tests were healthy. (Optimistic)
- Worst Case: Assume all missing tests were unhealthy. (Pessimistic)
- Analogy: Like guessing a student's final grade by assuming they aced all the tests they missed, or failed them all.
The "Missingness as a Category" Method:
Instead of guessing, they treated "missing a test" as its own unique category.
- Analogy: Instead of saying "Student A failed Math," you say "Student A has a 'Missing Math' status."
The "Pattern Submodels" Method (The Custom Groups):
This was the most complex strategy. They noticed that patients often miss the same combination of tests (e.g., if you miss the Homocysteine test, you probably also missed the C-Reactive Protein test). They grouped patients by these specific "missing patterns" and built a tiny, custom prediction model for each group.
- Analogy: Instead of one teacher grading the whole class, you have a different teacher for every group of students who missed the same specific tests.

The Results: What Worked Best?

The researchers tested these strategies to see which one best predicted who would be hospitalized.

Simple Math Beat Fancy AI: Surprisingly, a basic statistical tool called Logistic Regression (like a standard calculator) worked better than Random Forest (a complex, high-tech AI that builds many decision trees).
- Why? The data was already simplified into "Healthy" or "Unhealthy" (yes/no). There wasn't enough complex, non-linear data for the fancy AI to find hidden patterns. It was like using a supercomputer to solve a simple addition problem; the calculator was just as good and faster.
The "Count" vs. "Yes/No" Surprise: They tried to predict the exact number of hospital visits (0, 1, 2, etc.), but it didn't work better than just predicting "Did they go to the hospital? Yes or No?"
- Takeaway: Knowing if someone will be hospitalized is easier to predict than knowing how many times they will go.
The "Summary Score" vs. "Individual Parts":
- Summary Scores: When they combined the ten markers into one single number (the ALI score), all the different ways of handling missing data performed almost the same. The "Proportion" method was slightly the best, but barely.
- Individual Parts: When they looked at each marker separately, the Pattern Submodels (the custom groups) performed the best on the specific data they had (AUC = 0.73).
- The Catch: When they tested these custom models on new data (cross-validation), they didn't hold up as well (AUC dropped to 0.63). It's like a student who memorized the answers to a specific practice test but struggled when the questions changed slightly.

The Conclusion

The paper concludes that:

Missing data is tricky: You can't just ignore it or guess randomly.
Keep it simple: For this specific type of health data, a simple statistical model worked better than complex machine learning.
The "Pattern" approach is promising but needs more practice: Grouping patients by their specific missing data patterns gave the best results in the test, but it needs more data to prove it works reliably in the real world.

The Bottom Line: The researchers successfully showed how to build a "Whole-Person Health Score" even when data is incomplete. They found that while complex methods look good on paper, simple, tailored approaches often work best for predicting hospital visits in this specific context. The next step they mention is simply trying to put this system into real hospital charts to help doctors make decisions.

Technical Summary: Predicting Hospitalization from a Whole-Person Health Score with Incomplete EHR Data

Problem Statement
The transition from reactive to proactive healthcare relies on identifying at-risk patients before crises occur. Electronic Health Records (EHR) offer vast data for this purpose, but they suffer from high rates of missing values due to clinical decision-making (e.g., tests ordered only when necessary) and access barriers. The Allostatic Load Index (ALI), a standardized measure of whole-person health calculated from ten biomarkers across cardiovascular, metabolic, and inflammatory systems, is a promising tool for preventative care. However, in observational EHR settings, not all patients undergo all ten required tests. Standard calculation of ALI (counting unhealthy biomarkers) implicitly treats missing components as "healthy," potentially misrepresenting patient risk. This study addresses the challenge of predicting inpatient hospitalization using an ALI derived from incomplete EHR data, evaluating how different statistical and machine learning strategies handle missing components.

Methodology
The study utilized a dataset of $n=1000$ patients (aged 18–65) from Atrium Health Wake Forest Baptist Hospital between March 2018 and March 2020. The outcome variables were inpatient hospitalizations during the study period, modeled as both a binary indicator ( $Y_B$ : at least one hospitalization) and a count ( $Y_C$ : number of hospitalizations). Covariates included age and biological sex.

The authors evaluated 32 pairings of modeling techniques and missing data handling strategies:

Modeling Techniques:
- Statistical Regression: Logistic regression (for binary outcomes) and Poisson/Zero-Inflated Poisson/Negative Binomial regression (for count outcomes).
- Machine Learning: Random Forest classifiers.
Missing Data Strategies:
- Summary Measures (Aggregate ALI):
  - Complete-case proportion: Calculating the proportion of non-missing components that are unhealthy, ignoring missing values.
  - Counts of unhealthy and missing: Using the count of unhealthy components and the count of missing components as separate predictors.
  - Best/Worst case imputation: Treating all missing components as healthy (best case) or unhealthy (worst case).
- Separate Components (Disaggregated):
  - Missingness as a category: Including indicators for whether each component is missing, healthy, or unhealthy.
  - Pattern submodels: Fitting distinct logistic regression models for each unique missing data pattern observed in the cohort (18 patterns identified), predicting outcomes based on the specific submodel matching the patient's missingness profile.

Performance was evaluated using Receiver Operating Characteristic (ROC) curves and the Area Under the Curve (AUC) on the full sample (optimistic) and via 5-fold cross-validation (generalizable).

Key Results

Outcome Modeling: Modeling hospitalization as a binary outcome performed comparably to or better than count-based models. Logistic regression generally outperformed Random Forests in summary measure models, while Random Forests showed marginal improvements in specific separate-component scenarios.
Summary Measures: All summary measure approaches (complete-case proportion, counts, imputation) performed similarly. The complete-case proportion ALI achieved the highest full-sample AUC of 0.64 (cross-validated AUC 0.63), but the improvement over other summary methods was negligible ( $\le 0.01$ ).
Separate Components vs. Summary: When ALI components were included separately, performance varied significantly by missing data strategy.
- Pattern Submodels: This approach yielded the highest full-sample AUC (0.73) using logistic regression. However, it did not cross-validate well (AUC dropped to 0.63), suggesting potential overfitting due to rare missing data patterns (some with as few as one patient).
- Random Forest with Best-Case Imputation: Among separate component models, this method achieved the highest cross-validated AUC (0.66).
Overall Predictive Ability: Predictive performance across all models was fair to low (AUC $\le 0.73$ ). The authors note that while pattern submodels showed promise in-sample, their transportability is questionable without larger sample sizes to stabilize the rare pattern submodels.

Significance and Claims
The paper claims that while summary measures (like the complete-case proportion) offer interpretability and ease of implementation in clinical charts, they provide only fair predictive power for hospitalization. The study demonstrates that tailoring models to specific missing data patterns (pattern submodels) can improve in-sample accuracy by leveraging the information contained in the missingness itself, which is often informative in EHR data (Missing Not At Random).

However, the authors remain modest regarding the immediate clinical utility of these complex models. They conclude that the pattern submodel approach, while theoretically superior in capturing granular risk, requires larger datasets to validate its stability and prevent overfitting. The primary contribution is the evaluation of these specific strategies for handling incomplete ALI data, suggesting that future implementations should consider disaggregating components and accounting for missingness patterns, but acknowledging that current predictive performance remains limited. The authors propose that the next step is EHR implementation to enable prediction at scale, supported by further research into data quality and measurement error correction.