Predicting individual hemoglobin abnormalities using longitudinal data in clinical practice

Maliheh Namazkhan, Karel Jan van Tuijn, Maurits Kaptein, Remco van HorssenBmc Medical Informatics Decis. mak 2025 · DOI

TL;DR

This paper introduces a novel approach using Generalized Additive Models (GAMs) to predict individual hemoglobin (Hb) levels over time, moving beyond population-based reference ranges. By focusing on personalized "normal" bounds, the model aims to enable earlier detection of potential abnormalities, leading to more proactive and efficient preventive healthcare.

What Problem Does This Paper Address?

Preventive medicine and personalized healthcare are increasingly emphasizing early detection and intervention to improve health outcomes. A key challenge in this domain is the reliance on population-based reference ranges for blood parameters like hemoglobin (Hb). These broad ranges often fail to account for significant inter-individual variability, meaning a value considered "normal" for the general population might actually represent a concerning deviation for a specific individual. This can lead to missed early warning signs or unnecessary medical interventions. The paper addresses this gap by proposing a method to predict an individual's Hb trajectory and identify deviations from their own established normal range, rather than a generic one.

Methods

The core of this study lies in the application of Generalized Additive Models (GAMs). GAMs are statistical models that allow for flexible, non-linear relationships between predictor variables and the response variable. This is particularly advantageous for modeling longitudinal data where patterns can be complex and dynamic.

Key aspects of their methodology include:

  • Longitudinal Data Modeling: GAMs are well-suited for time-series data, allowing them to capture temporal trends and variations in Hb levels.
  • Personalized Thresholds: Instead of population-based ranges, the study establishes individualized "normal" ranges defined as mean ± 2.5 times the standard deviation of a patient's historical Hb measurements.
  • Predictor Variables: The model incorporates several predictors, including:
  • Average Hb: The patient's historical average Hb value.
  • Time since first Hb measurement: To capture temporal dynamics.
  • Patient Age and Gender: To account for demographic influences.
  • Random effect for patient ID: To capture unobserved individual variability.
  • Model Selection: Akaike's Information Criterion (AIC) and Bayesian Information Criterion (BIC) were used for model selection.
  • Performance Evaluation: Metrics like Mean Squared Prediction Error (MSPE), accuracy, precision, recall, and F1 score were used to assess predictive performance.

The novelty lies in the specific application of GAMs for this personalized Hb prediction task and the focus on identifying deviations from individualized normal bounds, moving beyond traditional statistical approaches.

Datasets

The study utilized a dataset comprising Hb measurements from 118,270 unique primary care patients spanning from 2012 to 2019.

  • Training/Development Data: A subset of 30,000 patients was used for computational feasibility.
  • Evaluation Data: The model was trained on all Hb measurements except the most recent one for each patient, and then evaluated on the most recent measurement to assess predictive performance on unseen data.
  • Sample Size for Analysis: A final sample of 56,453 patients (33,432 females, 23,021 males) was used after applying stringent inclusion criteria (e.g., Hb within a certain range, 3-100 measurements per patient).

No new benchmarks were introduced; the focus was on applying existing statistical methods to a novel clinical prediction problem.

Key Insights and Ideas

  • Individualized Baselines are Crucial: The study strongly reinforces the idea that relying solely on population-based reference ranges for Hb is insufficient. Individual variability is significant, and personalized thresholds are key to detecting meaningful deviations.
  • GAMs Offer Superior Flexibility: GAMs proved to be more adept than traditional linear models or even mixed-effects models in capturing the complex, non-linear relationships and temporal dynamics inherent in longitudinal Hb data.
  • Predictive Accuracy is Achievable: The model demonstrated a high degree of accuracy in predicting whether a patient's Hb levels would remain within their individual normal bounds, suggesting practical utility.
  • Preventive Potential: The ability to predict potential abnormalities before they become clinically significant opens doors for proactive interventions, potentially reducing unnecessary blood draws and improving patient outcomes.

Key Results

  • The model accurately predicted whether patients' Hb levels would stay within their individual "normal" bounds or deviate from them in 88.47% of cases.
  • The overall prediction accuracy was approximately 88.47%.
  • Precision was 88.4%, and recall (sensitivity) was exceptionally high at 99.95%, indicating the model's success in identifying true positives.
  • The Mean Squared Prediction Error (MSPE) was 0.48 mmol/L, suggesting that predictions deviate from true values by an average of 0.48 units.
  • Approximately 60.12% of predictions were within an acceptable error threshold of ± 0.5 mmol/L.
  • The study found that while random effects for patient ID contributed to capturing variability, the smooth term for average Hb effectively captured most of this individual-specific variation.

Limitations and Future Work

  • Monitoring System Development: The authors acknowledge a lack of a fully developed method to monitor deviations from individualized thresholds over time and suggest this as a potential area for future research.
  • Model Complexity vs. Performance: While a more sophisticated model with interaction effects showed a lower AIC, it did not translate to better predictive accuracy (higher MSPE). This suggests that balancing model complexity with practical predictive performance is crucial.
  • Further Predictors: Future research could explore additional factors that uniquely influence Hb levels, such as weight, dietary habits, physical activity, environmental exposures, and medical history.
  • Alternative Probabilistic Approaches: Investigating alternative probabilistic methods for defining individualized reference ranges (e.g., percentile-based thresholds, Bayesian credible intervals) is suggested.
  • Integration with Alerting Mechanisms: Developing a robust monitoring system that integrates the prediction model with automated alerting mechanisms is a key future direction to facilitate early intervention.

Summary updated

BibTeX
@article{namazkhan2025predicting,
  author = {Maliheh Namazkhan and Karel Jan van Tuijn and Maurits Kaptein and Remco van Horssen},
  title = {{Predicting individual hemoglobin abnormalities using longitudinal data in clinical practice}},
  journal = {Bmc Medical Informatics Decis. mak},
  year = {2025},
  doi = {10.1186/s12911-025-03085-6},
}
Figure