JMIR Publications

ABSTRACT

Background:

Neonatal disease and its outcomes are important indicators for responsive health care system and encompasses the effect of socio-economic and environmental factors on new-borns and mothers. Ethiopia is working to achieve the SDG target for the reduction of 12 or less per 1000 birth by 2030 and 21 per 1000 livebirths by 2025 as part of second Ethiopian Health Sector Transformation Plan.

Objective:

This study aimed to compare the performance of classical time-series models with that of deep learning models and to forecast the neonatal mortality rate in Ethiopia to verify whether Ethiopia will achieve national and international targets.

Methods:

Data were extracted from the official World Bank database. Classical time-series models, such as autoregressive integrated moving average (ARIMA) and double exponential smoothing, and neural network-based models, such as MLP, CNN, and LSTM, have been applied to forecast neonatal mortality rates from 2021 to 2030 in Ethiopia. During model building, the first 21 years of data (from 1990 to 2010) were used for training, and the remaining 10 years of data were used to test model performance. Commonly applicable measures were used to evaluate the predictive performance of the forecasting methods. Finally, the best model was used to forecast the neonatal mortality rate over the next 10 years from 2021 to 2030, with a 95% prediction interval.

Results:

The results showed that the double exponential smoothing model was the best, with a maximum R2 of 99.94% and minimum MAPE and RMSE of 0.0015 and 0.0748, respectively. The worst performing among the five models was the CNN, with an R2 of 93.71% and a maximum RMSE of 0.7903. Neonatal mortality in Ethiopia is forecasted to be 23.2 (PI: 22.2, 24.4) per 1000 live births in 2025 and 19.8 (PI: 17.1, 22.8) per 1000 live births in 2030.

Conclusions:

This study revealed that national and international targets for neonatal mortality cannot be realised if the current trend continues.