Improving the value of population health data for health policy and decision-making using machine learning algorithms in EQ-5D-5L index estimation – Scientific Reports

How much can we squeeze out of the health data we already collect? A new study puts that question to the test, showing that machine learning can accurately estimate EQ-5D-5L health utility scores using standard sociodemographic information and the Minimum European Health Module (MEHM) from seven large population surveys (N = 9,324). The results point to a practical way to generate health-economic inputs when direct EQ-5D-5L collection isn’t feasible—while also flagging pitfalls that analysts should avoid.

The big picture

EQ-5D-5L is a cornerstone measure for health technology assessment and cost-utility analysis, but it’s not always captured in routine datasets. This study tested whether common survey variables—age, sex, education, employment, and self-reported health via the MEHM—can stand in for directly measured EQ-5D-5L indices. Using 14 different machine learning models evaluated across five research scenarios with a newly developed performance metric (the G score), the authors show that high-quality estimates are possible, especially when MEHM information is included and data imputation is avoided.

How the study was done

• Data: Seven extensive population surveys totaling 9,324 participants, combining sociodemographic variables with MEHM responses.
• Objective: Estimate patient-level EQ-5D-5L index values from routinely collected variables.
• Models: Fourteen machine learning approaches compared across five scenarios (e.g., with/without MEHM, with/without imputation).
• Evaluation: A recently developed G score used to rank models consistently across scenarios.

What the models revealed

• Overall winner: AdaBoost delivered the strongest performance, achieving the best mean rank across scenarios (2.87).
• Close contenders: Multilayer Perceptron (MLP) followed closely (mean rank 2.94), with XGBoost also performing strongly (mean rank 3.60).
• Best-case setup: AdaBoost achieved its top result when both sociodemographic and MEHM data were included and no imputation was applied (G = 0.955).
• Data matters: When the estimation used only sociodemographic variables (no MEHM), performance dropped (G = 0.871).

The takeaway is clear: adding MEHM considerably improves EQ-5D-5L estimation, and the choice of preprocessing—especially whether to impute missing data—can materially impact model performance.

Why this matters for policy and planning

• Filling evidence gaps: Health-economic evaluations often stall when EQ-5D-5L hasn’t been collected. These methods can populate analyses using variables already captured in official surveys and registries.
• Faster, cheaper insight: When primary data collection is impractical or too costly, ML-based estimates provide a pragmatic alternative to inform budget impact assessments, scenario planning, and equity analyses.
• Better use of existing infrastructures: Many national and regional datasets already include MEHM; this approach leverages that infrastructure to generate utility values without new fieldwork.

A note of caution: imputation can backfire

The study highlights a counterintuitive finding: some forms of data imputation can degrade model performance. In the top-performing setting, the authors avoided imputation altogether. For analysts, that means:

• Test preprocessing choices explicitly—don’t assume imputation helps.
• Conduct sensitivity analyses across multiple missing-data strategies.
• Document how imputation affects model accuracy and calibration as measured by the chosen performance metric.

Limits and responsible use

Even strong ML models are not a perfect substitute for real EQ-5D-5L data. Estimates are inherently tied to the variables included and to the population and surveys used for training. Transportability to other settings should be tested, and estimates should be presented with appropriate uncertainty and caveats. Whenever feasible, direct EQ-5D-5L data collection remains the gold standard—particularly for high-stakes reimbursement decisions.

Key numbers at a glance

• Datasets: Seven population surveys, N = 9,324.
• Models compared: 14 machine learning algorithms, assessed over five scenarios.
• Top performers by mean rank: AdaBoost (2.87), MLP (2.94), XGBoost (3.60).
• Best G score: 0.955 with AdaBoost when using both sociodemographics and MEHM and no imputation.
• Reduced performance without MEHM: G = 0.871 using sociodemographics alone.

What’s next

For health systems and statistical agencies, the path forward includes curating standardized pipelines that:

• Prioritize inclusion of MEHM or similar self-reported health modules in routine surveys.
• Benchmark multiple ML models, including AdaBoost, MLP, and XGBoost, against local data.
• Report performance using transparent, comparable metrics (such as the G score).
• Provide open, reproducible code for validation and policy review.

Bottom line

This study shows that EQ-5D-5L utility scores can be estimated reliably from widely available survey variables, especially when MEHM data are present and imputation is handled with care. That unlocks practical value from existing datasets for health policy and decision-making, while reinforcing a critical principle: when you can collect EQ-5D-5L directly, you should.