A Multitask VAE for Time Series Preprocessing and Prediction of Blood Glucose Level

Data preprocessing is a critical part of time series data analysis. Data from connected medical devices often have missing or abnormal values during acquisition. Handling such situations requires additional assumptions and domain knowledge. This can be time-consuming, and can introduce a significant bias affecting predictive model accuracy and thus, medical interpretation. To overcome this issue, we propose a new deep learning model to mitigate the preprocessing assumptions. The model architecture relies on a variational auto-encoder (VAE) to produce a preprocessing latent space, and a recurrent VAE to preserve the temporal dynamics of the data. We demonstrate the effectiveness of such an architecture on telemonitoring data to forecast glucose-level of diabetic patients. Our results show an improvement in terms of accuracy with respect of existing state-of-the-art methods and architectures.

💡 Research Summary

This paper addresses a critical bottleneck in continuous glucose monitoring (CGM) data analysis: the need for extensive preprocessing to handle missing or abnormal values before forecasting blood glucose levels. Traditional pipelines separate imputation from prediction, relying heavily on domain expertise and ad‑hoc assumptions that can introduce bias and degrade model performance. To overcome these limitations, the authors propose a unified multitask architecture that combines a variational auto‑encoder (VAE) with recurrent neural networks (RNNs) to simultaneously perform data preprocessing (imputation) and long‑term glucose prediction.

The core of the method is a VAE that maps each multivariate time‑step into a latent Gaussian distribution parameterized by a mean vector μ and a log‑variance vector log σ². The encoder and decoder are built on top of RNN cells (LSTM or GRU), allowing the model to capture temporal dependencies while learning a compact latent representation. By conditioning the latent variable at each time step on the previous hidden state (as in a Variational Recurrent Neural Network, VRNN), the architecture preserves sequential dynamics and can infer missing observations from surrounding context.

Training optimizes a composite loss: (1) a reconstruction loss (L_reco) measured by mean‑squared error (MSE) between the original input and its reconstruction, (2) a prediction loss (L_pred) also measured by MSE but comparing the model’s forecast for a future horizon (30 minutes or 1 hour) with the true glucose values, and (3) a Kullback‑Leibler divergence term (L_KL) that regularizes the latent distribution toward a standard normal prior. The total loss is L_total = α L_reco + β L_pred + γ L_KL, where α, β, γ balance the importance of imputation, forecasting, and regularization.

The authors evaluate the approach on the public OhioT1DM 2018 dataset, which contains eight weeks of CGM, insulin dosing, and self‑reported meals for six type‑1 diabetes patients. Data are split 80 %/20 % for training and testing, and models are trained for 20 epochs. Baselines include simple statistical methods (forward‑fill, linear trend, ARIMA) and a suite of RNN variants (LSTM, Bi‑LSTM, GRU, Bi‑GRU). Performance is reported using Root Mean Squared Error (RMSE), Normalized Mean Absolute Percentage Error (nMAPE), and Mean Absolute Percentage Error (MAPE) for two prediction horizons: 30 minutes (6 steps) and 1 hour (12 steps).

Results show that the proposed VAE‑GRU model achieves the lowest error metrics across both horizons. For the 30‑minute horizon, VAE‑GRU attains RMSE = 26.93 ± 7.19, nMAPE = 11.73 ± 3.75, and MAPE = 11.99 ± 3.11, outperforming the best RNN baseline (Bi‑GRU) by roughly 10 % relative RMSE reduction. For the 1‑hour horizon, VAE‑GRU records RMSE = 39.92 ± 9.73, nMAPE = 18.42 ± 5.64, and MAPE = 19.16 ± 4.59, again edging out Bi‑GRU and matching or surpassing the VAE‑LSTM variant.

Beyond statistical metrics, the authors assess clinical relevance using the Clarke Error Grid, which categorizes predictions into zones A–E based on potential therapeutic risk. VAE‑RNN models place a higher proportion of predictions in zone A (clinically accurate) compared with all baselines, indicating that the integrated approach not only improves numeric accuracy but also reduces the likelihood of dangerous treatment decisions.

Key contributions of the work are: (1) an end‑to‑end multitask framework that eliminates the need for separate imputation pipelines, (2) a latent‑space driven imputation mechanism that leverages temporal context to fill gaps, and (3) the incorporation of a temporal attention mechanism within the VAE to better capture long‑range dependencies in glucose dynamics. Limitations include increased model complexity, higher computational cost, and the need for patient‑specific fine‑tuning. Future directions suggested are multi‑patient transfer learning, integration of exogenous variables such as meals and physical activity, and model compression for real‑time deployment on wearable devices. Overall, the paper demonstrates that a carefully designed VAE‑RNN architecture can substantially improve both the preprocessing and forecasting stages of CGM data analysis, offering a more robust and clinically safe tool for diabetes management.