PaperlayerView exact revision on arXiv

Source-linked AI summary

Predicting Blood Glucose with an LSTM and Bi-LSTM Based Deep Neural Network

Qingnan Sun, Marko V. Jankovic, Lia Bally, Stavroula G. Mougiakakou

arXiv:1809.03817v1cs.LGq-bio.QMstat.ML

TL;DR

The paper addresses predicting future blood glucose so patients or control algorithms can act proactively against high or low levels. It evaluates a glucose-based LSTM and bidirectional LSTM network against ARIMA and SVR, and reports better performance across prediction horizons and evaluation criteria.

  • Problem

    Future blood glucose prediction could let patients or control algorithms act proactively to avoid hyperglycaemia and hypoglycaemia.

    Source: §I · p. 1

  • Method

    A deep network using LSTM and bidirectional LSTM layers predicts blood glucose from CGM measurements, with pre-training using in silico and real-patient data.

    Source: §I · p. 2, C. Prediction model and training setup · p. 2, §V · p. 4

  • Results

    Across prediction horizons, the LSTM network outperformed ARIMA and SVR, reducing RMSE and time lag while increasing correlation coefficient and Fit.

    Source: §V · p. 4, §IV · p. 4

  • Takeaways & Limitations

    Because it needs only CGM measurements, the method could be used by patients using oral drugs, insulin pens, or a CSII pump.

    Source: §V · p. 4

    The preliminary study used only blood glucose data as prediction features; the authors identify adding features for specific user groups as future work.

    Source: §I · p. 2, §V · p. 4

Abstract

from arXiv · show

A deep learning network was used to predict future blood glucose levels, as this can permit diabetes patients to take action before imminent hyperglycaemia and hypoglycaemia. A sequential model with one long-short-term memory (LSTM) layer, one bidirectional LSTM layer and several fully connected layers was used to predict blood glucose levels for different prediction horizons. The method was trained and tested on 26 datasets from 20 real patients. The proposed network outperforms the baseline methods in terms of all evaluation criteria.

I. INTRODUCTION

Continuous glucose monitoring makes future blood glucose prediction a potential basis for proactive responses to high and low glucose levels. This study uses glucose-only inputs to test an LSTM and bidirectional LSTM predictor against ARIMA and SVR.

  • Diabetes-related hyperglycaemia is associated with organ damage, while hypoglycaemia can lead to unconsciousness, coma, or death.The introduction identifies preventing both extremes as important for keeping blood glucose within a healthy range.Source: §I · p. 1
  • CGM measures blood glucose every five minutes, making it possible to predict future levels so patients or control algorithms can act proactively.Source: §I · p. 1
  • Earlier work explored statistical and machine-learning predictors, while classical RNNs face limited long-term dependency learning because of vanishing or exploding gradients.LSTM addresses this challenge with memory cells and a forget gate; bidirectional LSTM can access past and future context.Source: §I · pp. 1–2
  • This study evaluates an LSTM and bidirectional LSTM deep network against ARIMA and SVR, using only blood glucose data as prediction features in this preliminary study.Source: §I · p. 2

II. METHODOLOGY

The methodology presents an LSTM network whose memory and gates address long-term dependencies, then details how its gates process and update the cell state.

A. LSTM network

  • LSTM addresses the long-term dependency problem in classical RNNs by introducing memory and a gate structure.Source: A. LSTM network · p. 2
  • The LSTM cell has input, forget, control, and output gates that regulate information transfer, memory retention, cell-state updates, and output generation.The input and forget gates use the previous hidden state and current input; the control gate updates the cell state, and the output gate generates the output and hidden vector.Source: A. LSTM network · p. 2

B. Bidirectional LSTM network

Bidirectional RNNs process information in both time directions; this work uses a bidirectional LSTM to access context from the past and future.

B. Bidirectional LSTM network

C. Prediction model and training setup

The prediction model combines LSTM, bidirectional LSTM, and fully connected layers, with pre-training and cross-validation used in its training setup.

C. Prediction model and training setup

III. DATASETS AND EVALUATION CRITERIA

The work uses both real-patient and in-silico datasets. Figure 3 presents the input and output dimensions of each layer.

  • Both real-patient and in-silico datasets were used.Source: §III · p. 2
  • Figure 3 presents the input and output dimensions of each layer.Source: Figure 3 · p. 3

A. Datasets from real patients

The real-patient data comprise CGM records from 20 adults with type 1 diabetes, preprocessed and selected into 26 qualifying sub-datasets.

  • The 20 adults with T1D had a mean age of 35 ± 14 years and diabetes duration of 17 ± 10 years.The datasets contained 6975 ± 1612 CGM measurements, sampled every five minutes.Source: A. Datasets from real patients · p. 3
  • Preprocessing addressed single outliers and measurement gaps, using linear interpolation for the former and splitting records with gaps into sub-datasets.Source: A. Datasets from real patients · p. 3
  • 26 sub-datasets met the minimum length of 1500 measurements, with lengths of 1791 ± 141; shorter datasets were merged for pre-training.The minimum represented at least five days of measurements.Source: A. Datasets from real patients · p. 3

B. Datasets from in silico trials

In-silico trials used the UVa/Padova T1D Simulator and virtual CGM data; the section also defines the glucose thresholds for hypo- and hyperglycaemia.

C. Evaluation criteria

Prediction performance was evaluated using RMSE, correlation coefficient, time lag, and fit. The criteria compare test datasets with predicted values.

  • The evaluation uses RMSE, correlation coefficient (CC), time lag (TL), and fit to compare test datasets with predictions.Source: C. Evaluation criteria · p. 3
  • RMSE is the square root of the mean squared error between target and predicted data; lower RMSE indicates better prediction.Source: C. Evaluation criteria · p. 3
  • CC represents the linear dependence between actual and predicted glucose values, with its calculation using standard deviations and covariance.Source: C. Evaluation criteria · p. 3
  • Time lag is the minimum shift yielding the highest correlation, while fit uses RMSE relative to the target's deviation from its mean; higher fit indicates better performance.Source: C. Evaluation criteria · p. 3

IV. EXPERIMENTAL RESULTS

Across 15-, 30-, 45-, and 60-minute prediction horizons, LSTM outperformed ARIMA and SVR on all reported evaluation criteria. The experiments also selected pre-train epochs by balancing performance and time cost.

IV. EXPERIMENTAL RESULTS

  • The experiments compared ARIMA, SVR, and LSTM at prediction horizons of 15, 30, 45, and 60 minutes.Source: §IV · p. 3
  • LSTM outperformed ARIMA and SVR across all prediction horizons on RMSE, time lag, correlation coefficient, and fit.Compared with both baselines, it lowered RMSE and time lag while increasing correlation coefficient and fit.Source: §IV · p. 4, Table II · p. 4
  • For PH=30 minutes, pre-train epoch settings such as 900, 1300, 1500, and 1700 performed well across all indices; 1300 was chosen to balance performance and time cost.The selected pre-train epoch counts for PH=15, 45, and 60 were 900, 1200, and 1800, respectively.Source: §IV · p. 4

V. CONCLUSION

The LSTM network outperformed ARIMA and SVR across prediction horizons, reducing RMSE and time lag while increasing correlation and fit. It was trained and tested on 26 real-patient datasets, with future work planned to add features and event alarms.

V. CONCLUSION

  • The network combined one LSTM layer, one bidirectional LSTM layer, and fully connected layers, and was trained and tested on 26 real datasets with more than five days of CGM data each.Pre-training used both in silico and real-patient data to generate a global model; the low training-epoch count was intended to enable rapid operation, including on mobile platforms.Source: §V · p. 4
  • Across all prediction horizons, the LSTM network outperformed ARIMA and SVR, reducing RMSE and time lag while increasing correlation coefficient and fit.Source: §V · p. 4
  • Table III presents LSTM prediction performance across pre-train epoch settings for the 30-minute prediction horizon.Source: Table III · p. 4
  • The method uses CGM measurements alone and could be used by patients on oral drugs, insulin pens, or CSII; planned work includes adding features for target groups and implementing event alarms.Source: §V · p. 4

Related papers

Loading 1809.03817v1…