Scheduled Machine Sampling (SMS): A Hybrid Approach for Glucose Forecasting
Issued Date
2024-01-01
Resource Type
Scopus ID
2-s2.0-85216518434
Journal Title
19th International Joint Symposium on Artificial Intelligence and Natural Language Processing, iSAI-NLP 2024
Rights Holder(s)
SCOPUS
Bibliographic Citation
19th International Joint Symposium on Artificial Intelligence and Natural Language Processing, iSAI-NLP 2024 (2024)
Suggested Citation
Kongpana K., Phienthrakul T. Scheduled Machine Sampling (SMS): A Hybrid Approach for Glucose Forecasting. 19th International Joint Symposium on Artificial Intelligence and Natural Language Processing, iSAI-NLP 2024 (2024). doi:10.1109/iSAI-NLP64410.2024.10799502 Retrieved from: https://repository.li.mahidol.ac.th/handle/20.500.14594/104251
Title
Scheduled Machine Sampling (SMS): A Hybrid Approach for Glucose Forecasting
Author(s)
Author's Affiliation
Corresponding Author(s)
Other Contributor(s)
Abstract
This paper introduces Scheduled Machine Sampling (SMS), a technique to improve Recurrent Neural Networks (RNNs) based time series forecasting by combining machine learning models with scheduled sampling. Unlike traditional Scheduled Sampling, which probabilistically blends one-step-ahead predictions with ground truth, SMS uses a blend of in-step predictions from a machine learning model and ground truths sampled from an exponential distribution with a fixed decay constant. Evaluated on the OhioT1DM dataset using a minimal feature set, the incorporation of SMS significantly improved forecasting accuracy, yielding average enhancements of 15.11% in Root Mean Square Error (RMSE) and 14.42% in Mean Absolute Error (MAE) across various models and sequence lengths, specifically for 15-minute, 30-minute, 1-hour, and 2-hour intervals. The resulting RMSE values were recorded at 10.6, 13.9, 21.5, and 32.5 mg/dl, while MAE values achieved were 10.06, 12.7, 19.0, and 28.3 mg/dl, respectively. Furthermore, models utilizing SMS exhibited only a minimal increase in inference time of 0.35 milliseconds per step on average, despite the added complexity, highlighting the effectiveness of incorporating machine-derived features as signals into RNNs.