Machine learning models trained on synthetic datasets of multiple sample sizes for the use of predicting blood pressure from clinical data in a national dataset

PLOS ONE · 2 authors, 2 centres

AI SUMMARY

FIDELITY 100%

POPULATIONAdults aged 18–70 years from the National Diet and Nutrition Survey (NDNS) Rolling Programme (2008–2019), excluding those on anti-hypertensive medications (n=2832 total, training n=2408, test n=424).

INTERVENTIONMachine learning models (Bayesian generalized linear regression, random forest, neural network) trained on synthetic datasets of varying sizes (n=2408, n=4816) and a combined real+synthetic dataset (n=4816).

COMPARISONSame machine learning models trained on the real training dataset (n=2408).

This summary was generated by AI from a single paper. It has not been reviewed by a clinician and is not clinical advice. Verify against the source before acting on it.

AGENT 01 · STRUCTURED SUMMARY

TRACE TO SOURCE

“

TL;DR

This study compared machine learning models trained on real versus synthetic data to predict blood pressure using only age, sex, weight, and height from a national dataset. Models trained on synthetic data performed equally well as those trained on real data, with mean absolute errors ranging from 8.12 to 8.33 mmHg. The findings suggest synthetic data could potentially replace real data for training predictive models, addressing data privacy concerns, though further validation is needed.

Full summary

2,919 CHARS

**Background:** The use of synthetic data in healthcare research is gaining interest as a way to increase data access and overcome privacy concerns. However, evidence that synthetic data can adequately train algorithms for use on real data is still limited. This study aimed to compare the performance of machine learning models trained on real versus synthetic data for predicting blood pressure from simple clinical variables.

**Methods:** Data were obtained from the National Diet and Nutrition Survey (NDNS) Rolling Programme (2008–2019). After exclusions (age <18 or >70, taking anti-hypertensives), 2832 participants remained. The dataset was split into training (85%, n=2408) and testing (15%, n=424). Eleven variables (age, sex, ethnicity, marital status, smoking status, socioeconomic status, total weekly screen time, frequency of takeaway meals, average nightly sleep duration, height, weight) were considered. Recursive feature elimination with tenfold cross-validation identified the optimal subset: age, sex, weight, and height. Missing data were imputed using the missForest package. Three model types were built using the caret package in R: Bayesian generalized linear regression (bayesglm), random forest (rf), and neural network (nn). Synthetic datasets were generated using the synthpop library: synthetic dataset A (n=2408), synthetic dataset B (n=4816), and a combined real+synthetic dataset (n=4816). All models were tested on the same real test set (n=424). Model performance was assessed using root mean squared error (RMSE), R-squared, mean absolute error (MAE), and residuals compared via Wilcoxon signed rank test.

**Key Results:** Synthetic datasets showed high fidelity with the real training data (propensity score mean-squared error values low; e.g., for age: pMSE=0.000225, S_pMSE=2.16 for synthetic A; pMSE=0.000011, S_pMSE=0.27 for synthetic B). Model performances were comparable across all training data types. For Bayesian linear regression: MAE ranged from 8.25 to 8.28; RMSE 10.45–10.47; R-squared 0.22. For neural network: MAE 8.12–8.19; RMSE 10.31–10.36; R-squared 0.24–0.25. For random forest: MAE 8.20–8.33; RMSE 10.49–10.71; R-squared 0.18–0.22. None of the Wilcoxon signed rank test p-values were <0.05, indicating no significant difference in residuals between models trained on real vs. synthetic data.

**Clinical Implications:** This study demonstrates that synthetic data can produce machine learning models with predictive accuracy equivalent to those trained on real data for estimating blood pressure using only age, sex, weight, and height. The minimalistic variable set enhances real-world applicability. However, the study used a non-representative sample and did not assess reidentification risk. Further research is needed to confirm generalizability and privacy protection before synthetic data can replace real patient data in clinical prediction modeling.

PICO

PPOPULATION

Adults aged 18–70 years from the National Diet and Nutrition Survey (NDNS) Rolling Programme (2008–2019), excluding those on anti-hypertensive medications (n=2832 total, training n=2408, test n=424).

IINTERVENTION

Machine learning models (Bayesian generalized linear regression, random forest, neural network) trained on synthetic datasets of varying sizes (n=2408, n=4816) and a combined real+synthetic dataset (n=4816).

OOUTCOME

Mean absolute error (MAE) and residuals of predicted mean arterial blood pressure on the same real test set.

STUDY TYPE

cross-sectional

SPECIALTY

cardiology

SUMMARISED BY

AI pipeline

FIDELITY CHECK

100% · A

ORIGINAL PAPER

PMC10019654

Machine learning models trained on synthetic datasets of multiple sample sizes for the use of predicting blood pressure from clinical data in a national dataset

PLOS ONE

READ THE FULL PAPER

ABSTRACT

Introduction The potential for synthetic data to act as a replacement for real data in research has attracted attention in recent months due to the prospect of increasing access to data and overcoming data privacy concerns when sharing data. The field of generative artificial intelligence and synthetic data is still early in its development, with a research gap evidencing that synthetic data can adequately be used to train algorithms that can be used on real data. This study compares the performance of a series machine learning models trained on real data and synthetic data, based on the National Diet and Nutrition Survey (NDNS). Methods Features identified to be potentially of relevance by directed acyclic graphs were isolated from the NDNS dataset and used to construct synthetic datasets and impute missing data. Recursive feature elimination identified only four variables needed to predict mean arterial blood pressure: age, sex, weight and height. Bayesian generalised linear regression, random forest and neural network models were constructed based on these four variables to predict blood pressure. Models were trained on the real data training set (n = 2408), a synthetic data training set (n = 2408) and larger synthetic data training set (n = 4816) and a combination of the real and synthetic data training set (n = 4816). The same test set (n = 424) was used for each model. Results Synthetic datasets demonstrated a high degree of fidelity with the real dataset. There was no significant difference between the performance of models trained on real, synthetic or combined datasets. Mean average error across all models and all training data ranged from 8.12 To 8.33. This indicates that synthetic data was capable of training equally accurate machine learning models as real data. Discussion Further research is needed on a variety of datasets to confirm the utility of synthetic data to replace the use of potentially identifiable patient data. There is also further urgent research needed into evidencing that synthetic data can truly protect patient privacy against adversarial attempts to re-identify real individuals from the synthetic dataset.

CITATION EXPORT

Anmol Arora, Ananya Arora. Machine learning models trained on synthetic datasets of multiple sample sizes for the use of predicting blood pressure from clinical data in a national dataset. PLOS ONE. (2023). https://doi.org/10.1371/journal.pone.0283094

AUTHORS

2 · 2 CENTRES

Anmol Arora, Ananya Arora

School of Clinical Medicine, University of Cambridge, Cambridge, United Kingdom
Jeonbuk National University, REPUBLIC OF KOREA

EDITOR / REVIEWER

Sathishkumar V E