Obesity is an established risk factor for non-communicable diseases such as type 2 diabetes mellitus, hypertension and cardiovascular disease (1–4). Lifestyle interventions involving weight loss have reduced the incidence of non-communicable diseases (5–7). A combined approach using dietary habits and physical activity is important for both short- and long-term.

Except for a few studies (8, 9), it has been shown that weight loss is difficult to achieve. The Trials of Hypertension Prevention (TOHP-II) showed that only 43% of participants lost more than 4 kg during a 6-month intensive behavioral intervention and that the drop-out rate was relatively high (10). Similarly, only 25% of participants lost at least 5% of their bodyweight between baseline and 12 months in an intervention conducted in primary care settings (11). These findings revealed the limitations of uniform treatment for weight control and that factors that influence weight vary substantially among individuals. Thus, an individualized approach to lifestyle interventions is essential for weight management.

Even when metabolic health was maintained over long periods, obesity was shown to remain a risk factor for cardiovascular disease (12). Thus, a simple and quick method to predict weight change over a few years could be helpful in weight management. In addition, not only identifying groups whose lifestyle profoundly impacts weight control but also clarifying factors that strongly influence an individual’s body weight is useful in clinical settings.

Machine learning, which can determine patterns and decision rules from data (13–16), is rapidly becoming used in clinical practice. The applications of machine learning for the early detection of diabetes, including its complications, and cancer that provided clear-cut diagnostic gold standards have been evaluated (17–24). However, little is known about the usefulness of machine learning to predict body weight over the long term, as well as to identify factors that influence body weight in individuals. Moreover, some issues involving the nature of the machine learning algorithm, which is often referred to as a black box model (13–16), could be a barrier to the practice of evidence-based medicine.

Heterogeneous mixture learning technology (HMLT) is a newly-developed method to automatically divide original data to increase the mining accuracy of patterns, trends, and rules in the data (25). Even when it is difficult to know the number of splits or what clues to use for splitting, it is possible to conduct appropriate splits at high speed based on characteristics of the new groups (25). Thus, this technology be expected to be superior to other machine learning models in predicting weight change. Although unexplained machine learning based weight prediction models for relatively short periods of time have been reported (26, 27), little is known about the effectiveness of explainable machine learning in predicting weight change over a relatively long period of time.

Therefore, we constructed a machine learning model to predict body weight over 3 years using medical checkup data. Using this information, we attempted to evaluate the ability of machine learning models for weight prediction and to identify factors that influence weight loss in individuals.

Characteristics of participants according to learning/validation data and sex are shown in Table 1. Mean age was 48 years and 67% were men. There were no significant differences between learning data and validation data except for height in men.

Figure 1 is a schema of branches of the predictive formulas that were produced from a data of information on participants using the HMLT model. In our case, HMLT automatically generated 5 formulas by a linear model based on a rule that maximizes the predictive capacity throughout the dataset. As a result, participants were classified into one of the following five categories, i.e., predictive formulas, according to the algorithm shown in the Figure 1. First, the population was stratified into two groups according to BMI <29.93 kg/m² and BMI ≥29.93 kg/m². Then, each group was stratified by BMI ≥ 23.44 kg/m² and BMI <23.44 kg/m² or sex (men/women). The group with BMI ≥23.44 kg/m² was further divided into two groups depending on whether they were ≤24 years of age or >24 years of age. Finally, people were divided into five groups.

Table 2 shows the predictive formula for each of the 5 categories according to the algorithm shown in Figure 1. Each of the coefficients is a coefficient of the linear model used to calculate weight at the predicted year by baseline weight and lifestyle changes. The coefficients were relatively high in formulas 1, 2, and 5. RMSE was 1.914, which reflects ability comparable to the multiple regression model and was 1.890 (p = 0.323; Supplementary Table S1).

Figure 2 shows the examples of trends of future body weight according to Figure 2A breakfast intake (yes or no) and Figure 2B walking speed (fast or not). The first two dots of each component show actual measured values for the first 2 years and the latter three dots show predicted values of body weight for the third year of the follow-up. People with healthy lifestyle habits, such as eating breakfast or fast walking, tended to experience weight reduction, while unfavorable lifestyle habits were shown to increase weight (Figure 2). Similar body weight trends were observed for other lifestyle factors.

We constructed a machine learning model to predict body weight over 3 years using medical checkup data. The machine learning model utilizing HMLT automatically generated five predictive formulas and the predictive ability of HMLT was similar to that of multiple regression. Also, HMLT automatically identified groups of people whose lifestyle habits had a profound impact on weight loss.

To our knowledge, this was the first trial to predict body weight over 3 years using artificial intelligence, which resolved the issue of the explainable predictions in a way that humans can understand. The machine learning model could predict weight change over a 3-year period. Those findings suggested that the machine learning model could contribute to an individualized approach to weight management in clinical settings.

Weight loss is a key factor in the prevention of non-communicable diseases. Body weight was shown to gradually increased before the onset of diabetes regardless of obesity level (35). The Diabetes Prevention Program showed that weight loss through lifestyle interventions was the dominant predictor of reduced diabetes incidence. In that study, there was a 16% reduction in risk for every kilogram of weight loss. It was also shown that weight reduction was associated with significant health benefits in Japanese (36, 37). However, weight loss is difficult to achieve, and about a half of participants failed to lose more than 4 kg during a 6-month intensive behavioral intervention with a relatively high drop-out rate (10). Although lifestyle factors were shown to significantly influence body weight (38, 39), the impact of each component of an individual’s lifestyle on weight loss was complicated and, moreover, varied widely among individuals. In our study, the coefficients were relatively high for formulas 1, 2, and 5 (Table 2). In other words, the influence of lifestyle on body weight was found to be large in people with a high BMI at baseline and in younger people with a low BMI in our cohort, which was not possible with conventional multiple regression analysis. An approach using machine learning may make it possible to identify those for whom particular lifestyle interventions in each regional (or ethnic) population would be appropriate, leading to a more effective use of limited medical resources.

A quick way to predict future weight change as well identify factors that influence weight change in individuals has been eagerly awaited in clinical practice. Although previous research provided several models to estimate individual weight change using dietary intake and/or physical activity (40, 41), few models have visualized details on which of these influence an individual’s weight change over the subsequent few years. Our findings made it possible to simulate an individual’s weight change accompanied by lifestyle changes at least over 3 years that was accompanied by lifestyle changes (Supplementary Figure S2). Thus, individualized instructions on lifestyle factors could possibly improve motivation for weight loss.

The HMLT automatically divides original data to increase the mining accuracy of patterns, trends, and rules in the data. Thus, a machine learning model can provide optimal predictive equations automatically. In this study, we constructed a system that automatically identified targets whose weight changed greatly after lifestyles factors had changed. Previous reviews showed that many factors influence body weight (38, 39), suggesting that the effectiveness of standardized treatment used in clinical settings among individuals. Our findings indicated that individualized guidance may help achieve weight loss efficiently in clinical practice.

Thus far, the most important criticism of artificial intelligence is its clinical context as a “black box” which cannot easily explain the reasons for and background of results learned by the model, which could be a fatal barrier for its current use in clinical practice and care, which is required to be based on evidence based medicine (42–44). Indeed, there are some studies that have used artificial intelligence to predict weight at a single point, but those models cannot be explained in a way that humans can understand or predict predicted body weight only once (26, 27). The HMLT is epoch-making in that it can be understood by providers and recipients of medical services. In fact, this method can construct predictive models from data quickly and automatically. In addition, large amounts of time are not required to create and verify complicated predictive formulas. Thus, the workload of medical staff would not only be reduced but benefits would accrue to patients.

Our study had several limitations. First, we could not obtain certain information, such as information on family members and socioeconomic status, that would affect body weight (3, 4). Second, we included only participants who underwent a medical examination. Third, the information on diet and exercise was assessed using a self-reported questionnaire. Fourth, the fact that the study population was exclusively ethnic Japanese could limit wider applicability of the results.

In conclusion, the machine learning model could predict weight change over a period of 3 years. Our model automatically identified groups whose lifestyles have a profound impact on weight loss and the factors that influence body weight in individuals. Those findings suggested the possibility that a machine learning model could support individualized guidance for weight management in clinical settings.

The datasets presented in this article are not readily available due to proprietary reasons. Queries regarding the datasets should be directed to kafujihara-dm@umin.ac.jp.

The studies involving human participants were reviewed and approved by Niigata University (2020-0003). The patients/participants provided their written informed consent to participate in this study.

HS: full access to all data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. KF, MY, CH, MI, and TY: study concept and design. KF and KK: acquisition of data. KF, HT, HN, YS, and KT: analysis and interpretation of data, drafting of the manuscript. KF, HN, SK, KK, and HS: critical revision of the manuscript for important intellectual content. SK, KK, and HS: study supervision. All authors contributed to the article and approved the submitted version.

This work was supported in part by the Japan Society for the Promotion of Science and the Ministry of Health, Labor and Welfare, Japan (19H04028, 21 K11569). The sponsor or funding organization had no role in the design or conduct of this research.

HT, HN, YS, and KT were employed by NEC Solution Innovators, Ltd.

The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.