JMIR Publications

ABSTRACT

Background:

Population-level forecasting of hypertension and cholesterol is important for cardiovascular disease prevention, health-system planning, and policy evaluation. However, many medical machine-learning forecasting studies rely on limited model comparisons, omit classical statistical baselines, and do not formally test whether apparent performance differences are statistically meaningful.

Objective:

This study aimed to benchmark classical statistical, classical machine-learning, deep-learning, and hybrid deep-learning models for short epidemiological time-series forecasting of age-standardized hypertension prevalence and mean total cholesterol in adult males.

Methods:

We evaluated 13 forecasting models using 54 annual observations of age-standardized hypertension prevalence and mean total cholesterol among adult males aged 30–79 years from 1970 to 2023. The models included classical time-series approaches, classical machine-learning models, single-architecture deep-learning models, and hybrid neural architectures. The hybrid models included a proposed convolutional neural network–bidirectional long short-term memory–attention model that integrates causal convolution, bidirectional recurrent representation learning, and multi-head self-attention using approximately 23,000 trainable parameters. All models were evaluated using a walk-forward validation protocol with an 8-year observed test window from 2016 to 2023. Forecast accuracy was assessed using six error metrics, including symmetric mean absolute percentage error and Theil’s inequality coefficient. Statistical differences in forecast accuracy were assessed using the Diebold–Mariano test with the Harvey–Leybourne–Newbold small-sample correction. Recursive forecasts were additionally evaluated at 1-, 2-, and 5-year horizons from an end-of-2015 anchor; 10-year forecasts were generated for projection analysis but were not used for observed-error comparison because they extended beyond the available empirical data.

Results:

Autoregressive integrated moving average achieved the lowest 1-step root mean square error for both outcomes, with values of 0.00209 for age-standardized hypertension prevalence and 0.05690 for mean total cholesterol. In corrected significance testing, only support vector regression performed significantly worse than autoregressive integrated moving average on both outcomes at the 5% significance level. The proposed convolutional neural network–bidirectional long short-term memory–attention model was not statistically distinguishable from autoregressive integrated moving average for either age-standardized hypertension prevalence or mean total cholesterol. Residual analysis showed approximately symmetric error distributions for autoregressive integrated moving average and the proposed hybrid model, whereas weaker extrapolative models showed more visible bias.

Conclusions:

For short annual epidemiological time series, model choice should be horizon dependent rather than based only on overall predictive ranking. Classical time-series models remained highly competitive, and the proposed hybrid model achieved comparable performance without statistically significant degradation relative to the strongest classical baseline. These findings support the routine inclusion of classical baselines, symmetric error metrics, walk-forward validation, and formal significance testing in medical forecasting studies.