Introduction

Broadly, there are two types of machine learning: generative and deductive. Deductive machine learning models are those which analyse datasets to yield inferences that can be applied when presented with novel data [1]. Generative models function by similarly analysing datasets, but with a view to produce new data resembling the real original data. In healthcare, deductive machine learning models have captured academic, clinical and media attention due to their increasing ability to gain population health insights from large datasets and use these to inform health policy. Generative artificial intelligence (AI) is a much newer phenomenon and the ability to create synthetic data has use cases including dataset augmentation and data privacy [2–4]. Generative AI models, including generative adversarial networks, function by creating synthetic datasets composed of uniquely generated ‘fake’ datapoints, which at an aggregate level maintain all the patterns of the original dataset. It is already being considered that AI-generated synthetic datasets may begin to be used in place of real datasets to train deductive machine learning models [5]. The main advantages of doing so are that anonymous synthetic data can be shared instead of real data and the size of the datasets can be artificially increased. These advantages rest upon the assumption that the synthetic data is representative of the original dataset and patterns are preserved. Several methods of assessing the fidelity of synthetic data have been proposed, including simplistic Turing tests, histogram analysis and comparing the outputs of research analysis with those from real data [6, 7]. The lack of uniformly accepted standards for assessing synthetic data is a major limitation to the field [8]. This is further complicated by the fact that data must maintain high fidelity with the original data, but must be sufficiently different that the datapoints are genuinely non-identifiable.

This study presents a series of machine learning models trained to predict blood pressure of individuals within the National Diet and Nutrition Study (NDNS), based on simple variables [9]. The prediction of blood pressure using machine learning is an area of research which has gained attention recent years, with models usually being trained on electrocardiogram (ECG) or photoplethysmography (PPG) data, though there are examples of models trained on risk factors [10, 11]. The results of machine learning models are compared with those produced by machine learning models trained on synthetic datasets, of varying sizes, representative of the National Diet and Nutrition Study. In each case, the data is tested on the same sample of real data, but is trained on either real or synthetic datapoints.

Methods

Study design and dataset

This was a cross-sectional retrospective machine-learning study. This study uses data from the National Diet and Nutrition Study Rolling Programme (NDNS) (2008–2019). Ethical approval for the NDNS was obtained from the Oxfordshire A Research Ethics Committee and the Cambridge South NRES Committee (Ref. No. 13/EE/0016). In this analysis, we use data on adults aged from 18 to 70 years, combined from the first eight years (2008–2019) of the NDNS to provide a sufficiently large sample size for analysis (Table 1). An upper age limit of 70 years was applied due to the likelihood of comorbidities affecting blood pressure in the elderly population. Participants who reported taking anti-hypertensive medications were also excluded from the analysis. Mean arterial pressure, which is the average arterial pressure throughout one cardiac cycle, was calculated as the outcome variable, using the following equation [12]:

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Descriptive statistics of the study population, grouped by age.

Only age, sex, height and weight were ultimately used in the machine learning models as predictive variables. All variables were, however, used for imputing missing data and constructing synthetic datasets.

https://doi.org/10.1371/journal.pone.0283094.t001

There were 55 datapoints missing socioeconomic status, 3 datapoints missing smoking status, 707 datapoints missing screen time, 1 datapoint missing takeaway meal frequency, 1 datapoint missing ethnicity, 96 datapoints missing sleep duration, 56 datapoints missing height and 71 datapoints missing weight. Missing datapoints were imputed by the multiple imputation package ‘missForest’ with all variables being used as predictors [13].

Synthetic data

Three synthetic datasets were constructed, each based on the original training dataset. No data from the test dataset was leaked into the generation of the synthetic datasets. The synthpop library in R was used to produce two datasets of sizes: n = 2408 and n = 4816 [16]. A third dataset (n = 4816) was also used in the analysis, which consisted of the real data training set and the synthetic dataset (n = 2408) combined. Fig 1 illustrates the basic demographic analyses of synthetic dataset A (n = 2408) datasets compared to the real training dataset in histogram form. Fig 2 illustrates the comparison between synthetic dataset B (n = 4816) and the real training dataset, with both synthetic datasets demonstrating high fidelity at an aggregate level. The same machine learning analysis described above was applied to each of the synthetic datasets with the same test dataset used for the testing of all models. Identical pre-processing was also applied to the synthetic datasets and the same variables were isolated for analysis.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Histogram comparison for each variable comparing the aggregate demographic characteristics of the real training dataset (n = 2408) against synthetic dataset A (n = 2408).

https://doi.org/10.1371/journal.pone.0283094.g001

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Histogram comparison for each variable comparing the aggregate demographic characteristics of the real training dataset (n = 2408) against synthetic dataset B (n = 4816).

https://doi.org/10.1371/journal.pone.0283094.g002

Results

Of the 15655 participants who took part in years 1–11 of the NDNS, 7697 had blood pressure values recorded. Of these, 3256 were aged between 18 and 70 years old. A further 410 participants were excluded for taking blood pressure lowering medication, leaving a sample size of 2832 for the study (Table 1).

Two synthetic datasets were constructed, of different sizes. Table A was the same size as the training dataset comprised of real data (n = 2408), Table B was double the size (n = 4816). The datasets were generally of high fidelity, when compared to the real training dataset with the aggregate descriptive analysis comparing each variable displayed in Table 2.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Statistical analysis comparing synthetic data tables to the real training dataset (n = 2408).

Presented are propensity score mean-squared-error and standardised ration of propensity score mean-squared error.

https://doi.org/10.1371/journal.pone.0283094.t002

The performances of the machine learning models are shown in Table 3. The three model types were comparable in their results. Wilcoxon signed rank test with continuity correction to compare the residuals of each model. It can be seen that for each model type, algorithms trained comparably, regardless of whether they were trained on real data, synthetic data or the augmented real dataset.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Results of the machine learning models, trained on real or synthetic datasets.

Each was tested on the same test dataset (real data). None of the p-values were <0.05.

https://doi.org/10.1371/journal.pone.0283094.t003

Discussion

Algorithms trained on synthetic data performed comparably to those trained on real data with no significant differences found between the two types. Furthermore, when comparing the descriptive statistics of the real and synthetic datasets, there was minimal difference between them at an aggregate level. Together, these findings support the hypothesis that synthetic data can be used to train machine learning algorithms with the intention that they be tested on real data. All algorithms were able to predict blood pressure with a mean absolute error of approximately 8.2mmHg. The only variables used to achieve this were: age, sex, weight and height. This feat was achieved based on the use of recursive feature elimination to ensure that unnecessary variables were not being including in the model. In predicting blood pressure, this study represents a minimalistic approach to variables used in training the machine learning models, which increases its applicability in the real world.

There have been previous research efforts to predict blood pressure using other elements of a patient’s healthcare record. This has often involved electrocardiogram (ECG) analysis and photoplethysmography (PPG) [17, 18]. Whilst capable of performing impressively, for example Wang et al achieved an MAE of 4.02 for systolic blood pressure, these are less applicable at a population level due to the need for physical devices and monitoring. Another study conducted by Nath et al produced an MAE of approximately 10mmHg using age, dual‐energy X‐ray absorptiometry (DXA) measured body composition parameters, BMI, and waist circumference [19]. Other studies have sought to simply predict whether an individual has hypertension or not, using similar methodologies. In these cases, accuracies of predictions have roughly been in the range of 80–90% for high performing models [20, 21]. Studies using population level variables such as in this study have typically focussed on predicting the presence or absence of hypertension. Indeed, this study represents one of the first studies, if not the first study to predict the actual blood pressure value using descriptive clinical data without the need for EEG or PPG monitoring [10]. Even of those studies which have used descriptive clinical data to predict the presence of hypertension, typically these have used specific variables relating to blood pressure, including doctor’s perception of their blood pressure and whether they have measured their blood pressure [21]. Predicting blood pressure is a potentially important research area as it could give rise to targeted public health measures to optimise one of the most well recognised predictors of cardiovascular disease and all-cause mortality [22, 23]. As well as directing influencing the management of hypertension at a population level, this realm of research has the potential to improve our understanding of the disease process underlying hypertension as we begin to understand which variables can be predictive of high blood pressure. By combining multimodal data, for example clinical, genetic and physiological datapoints we may be able to yield more accurate predictions in the future [24]. The ability of synthetic data to inflate sample sizes in training data is an area for future research. In this study, there was no significant difference between the performance of models trained on smaller (n = 2408) or larger (n = 4816) datasets. This suggests that the real data sample size of 2408 was sufficient to train the algorithms. Future research could attempt to train algorithms on smaller sample sizes and assess, in granular detail, the relationship of synthetic data sample size with model accuracy.

Research into synthetic data is still in its infancy, with studies beginning to emerge suggesting the potential of synthetic data to train machine learning models to produce comparable results to the training on real data. Although this study used data that was already publicly available, it has been suggested that the same method of synthetic data generation used here could be used to help researchers release open-access datasets in a synthetic version of the actual confidential data used for studies [25]. Data confidentiality and ownership is an important ethical barrier to the implementation of artificial intelligence in healthcare, with synthetic data emerging as a potential solution [26]. A practical example if the use of synthetic data to replicate the results of a published stage III colon cancer trial secondary analysis, with high concordance between the results of models based on real and synthetic data [27]. Generative forms of artificial intelligence have also been used to create other forms of synthetic data, for example time-series data within electroencephalogram (EEG) signals and to selectively generate fundus photos from underrepresented groups to re-balance retinal imaging data [28, 29]. However, despite the well-recognised advantages of synthetic data, concerns are beginning to materialise. This includes the risk of data being used maliciously, or as a means of bypassing data protection legislation [30]. More broadly, there are ethical implications involved with this line of research. For example, if insurance providers are able to predict a patient’s blood pressure or other health characteristics, they could use this information to adjust premiums.