Spain

In this study we investigate the need for training future blood glucose level prediction models at the individual level (i.e. per patient). Specifically, we train various model classes: linear models, feed-forward neural networks, recurrent neural networks, and recurrent neural networks incorporating attention mechanisms, to predict future blood glucose levels using varying time series history lengths and data sources. We also compare methods of handling missing time series data during training. We found that relatively short history lengths provided the best results: a 30 minute history length proved optimal in our experiments. We observed long short-term memory (LSTM) networks performed better than linear and feed-forward neural networks, and that including an attention mechanism in the LSTM model further improved performance, even when processing sequences with relatively short length. We observed models trained using all of the available data outperformed those trained at the individual level. We also observed models trained using all of the available data, except for the data contributed by a given patient, were as effective at predicting the patient&#39;s future blood glucose levels as models trained using all of the available data. These models also significantly outperformed models trained using the patient&#39;s data only. Finally, we found that including sequences with missing values during training produced models that were more robust to missing values.

ECAI 2020

Experiments in non-personalised future blood glucose level prediction

ECAI2020 featured a full programme of technical papers, as well as many AI-related events of interest for researchers, students and all attendants who are interested on contemporary AI, including:




**An extensive programme of workshops, tutorials, and invited speakers**

**The Prestigious Applications of Intelligent Systems conference (PAIS)**

**The Frontiers in AI track sessions**

**The Lab To Market event**

**The Women in AI time**

**The EU Challenges forum**

… and many more!

ECAI 2020 featured a full programme of technical papers, as well as many AI-related events of interest for researchers, students and all attendants who are interested on contemporary AI.

In this study we investigate the need for training future blood glucose level prediction models at the individual level (i.e. per patient). Specifically, we train various model classes: linear models, feed-forward neural networks, recurrent neural networks, and recurrent neural networks incorporating attention mechanisms, to predict future blood glucose levels using varying time series history lengths and data sources. We also compare methods of handling missing time series data during training. We found that relatively short history lengths provided the best results: a 30 minute history length proved optimal in our experiments. We observed long short-term memory (LSTM) networks performed better than linear and feed-forward neural networks, and that including an attention mechanism in the LSTM model further improved performance, even when processing sequences with relatively short length. We observed models trained using all of the available data outperformed those trained at the individual level. We also observed models trained using all of the available data, except for the data contributed by a given patient, were as effective at predicting the patient's future blood glucose levels as models trained using all of the available data. These models also significantly outperformed models trained using the patient's data only. Finally, we found that including sequences with missing values during training produced models that were more robust to missing values.

technical paper

This work suggests two methods—both relying on stacked regression and data fusion of CGM and activity—to predict the blood glucose level of patients with type 1 diabetes. Method 1 uses histories of CGM data appended with the average of activity data in the same histories to train three base regressions: a multilayer perceptron, a long short- term memory, and a partial least squares regression. In Method 2, histories of CGM and activity data are used separately to train the same base regressions. In both methods, the predictions from the base regressions are used as features to create a combined model. This model is then used to make the final predictions. The results obtained show the effectiveness of both methods. Method 1 provides slightly better results.

Data Fusion of Activity and CGM for Predicting Blood Glucose Levels

In the presentation, we explained the usage of Seq-2-Seq models for blood glucose levels prediction and explored the possibility of using Transfer Learning for the same purpose.  We provided a detailed explanation of the results from the corresponding paper.

Blood Glucose Level Prediction as TIme-Series Modeling using Sequence-to-Sequence Neural Networks

"Glucose control is an essential requirement in primary therapy for diabetes management. Digital approaches to maintaining tight glycaemic control, such as clinical decision support systems and artificial pancreas systems rely on continuous glucose monitoring devices and self-reported data, which is usually improved through glucose forecasting. In this work, we develop a multitask approach using
convolutional recurrent neural networks (MTCRNN) to provide short-term forecasts using the OhioT1DM dataset which comprises 12 participants. We obtain the following results - 30 min: 19.79 +/- 0.06 mg/dL (RMSE); 13.62 +/- 0.05 mg/dL (MAE) and 60 min: 33.73 +/- 0.24 mg/dL (RMSE); 24.54 +/- 0.15 mg/dL (MAE). Multitask learning facilitates an approach that allows for learning with the data from all available subjects, thereby overcoming the common challenge of insufficient individual datasets while learning appropriate individual models for each participant."

Personalised Glucose Prediction via Deep Multitask Networks

Control of blood glucose (BG) levels is essential for diabetes management, especially for long term health improvement. Predicting both hypoglycemic events (BG < 70 mg/dl) and hyperglycemic events (BG > 180 mg/dl) is essential in helping diabetics control their long term health. In this paper, we attempt to forecast future blood glucose levels, as well as analyze the efficiency of detecting both hypoglycemic events and hyperglycemic events. We do so by comparing Auto-Regressive Integrated Moving-Average, Vector Auto-Regression, Kalman Filter, Unscented Kalman Filter, Ordinary Least Squares, Support Vector Machines, Random Forests, Gradient Boosted Trees, XGBoosted Trees, Adaptive Neuro-Fuzzy Inference System (ANFIS), and Multi-Layer Perceptron in terms of Root Mean Squared Error, Mean Absolute Error, Coefficient of Determination, Matthews Correlation Coefficient, and Clarke Error Grid to compare their effectiveness in predicting future blood glucose levels, as well as predicting both hypoglycemic and hyperglycemic events.

Comparison of Forecasting Algorithms for Type 1 Diabetic Glucose Prediction on 30 and 60-Minute Prediction Horizons

People with Type 1 Diabetes must take great care to monitor their blood glucose levels and take actions to keep them from getting either too high or too low. Having a meal will raise blood glucose levels; however, the number of carbohydrates that should be consumed to reach a desired blood glucose level depends on the recent history of several factors, such as boluses and other meals.  We demonstrated the feasibility of a generic neural architecture for making recommendations for the number of carbs a meal should consist of in order to reach a specified target blood glucose level.

A General Neural Architecture for Carbohydrate and Bolus Recommendations in Type 1 Diabetes Managment

"Planning for unknown environments presents a number of technical challenges. The planner
must ensure robustness to unknown phenomena and manage unpredictable variation in
execution, all while operating in a capacity that maximizes its objective. Productivity in the
face of these challenges requires an integrated approach to planning and execution that is
capable of accomplishing goals, reacting to variation, and maximizing overall utility. We examine
this problem in the context of a Europa Lander concept mission. We model the problem as
a hierarchical task network, framing it as a utility maximization problem constrained on a
depletable energy resource. We propose a planning and execution framework that responds to
feedback using three techniques: (1) flexible execution, (2) periodic replanning, and (3) online
model parameter and utility updates. The efficacy of each of these techniques is examined
through simulation of a Europa Lander concept mission, showing higher utility achievement
compared to baseline approaches. We demonstrate that an integrated approach to planning
and execution that is grounded in replanning, utility maximization, and model parameter
updates will be an enabling technology for future tightly-constrained planetary surface
missions."

Using Flexible Execution, Replanning, and Model Parameter Updates to Address Environmental Uncertainty for a Planetary Lander

"Mars Express (MEX) is a spacecraft operated by the European Space Agency (ESA) and orbiting Mars since the beginning of 2004. It has provided a wealth of scientific data such as evidence of the presence of water on  and  below  the  surface  of  the  planet,  three-dimensional renders  of  the  surface,  as  well  as  data  about  the  chemical composition of the Martian atmosphere.  The  instruments  and on-board  equipment are  kept  within  their  operating temperature ranges by an autonomous thermal control system composed of 33 heater lines. The thermal system consumes a significant amount of the total available electric power, leaving only a fraction to be used for science operations. 

Within the project GalaxAI, we use state-of-the-art explainable machine learning methods for structured output prediction to address the challenge of optimal use of the available energy. We construct accurate models of the MEX’s thermal power consumption under a variety of operating conditions, such as gyro-less flying and limited data availability. Besides good predictive performance, these methods also facilitate explainability and understanding of the provided estimates. The predictions and the explanations are visualized through a specialized dashboard, facilitating the user interaction with the learned models, and allowing for planning further operations of the spacecraft by the mission planning and operations team.
"

GalaxAI: Machine Learning for Spacecraft Operations

Delayed and intermittent communication with rovers in long-distance missions remains a key challenge for teleoperation, as it affects critically either the liveness or the safety of the rover. This position paper proposes the use of the newly-introduced human-machine interaction paradigm of Machine Coaching, as a middle-ground approach between a fully remotely-controlled and a fully autonomously-operated rover. According to this paradigm, the rover makes decisions and takes actions based on a policy, under the in-situ (even if delayed) supervision of a human coach. Whenever the human coach observes an unexpected decision made by the rover, the former enquires about the reasoning of the latter in making that decision, and then provides feedback on why a possibly different decision (including that of waiting to receive instructions from a remote operator) would have been preferable. This particular protocol, through the bilateral exchange of explanations between the rover and the human coach, allows the rover to provably and efficiently converge towards a policy that is acceptable to the human coach, while also endowing the rover with conditional autonomy in those situations where its current policy offers, in the human coach's expert opinion, an appropriate balance between liveness and safety. The rover's initial policy can be designed either in a simulated environment during the rover's development, or it can be based on early information on the remote environment transmitted once the rover is deployed. 

Rover Teleoperation through Machine Coaching

"The dynamics of space objects in Low Earth Orbit are strongly determined by the effects of atmospheric drag. This complex interaction is considered one of the most dominant sources of uncertainty at these altitudes, and is characterised in part by the atmospheric density, a quantity highly correlated to space weather. Current atmosphere models typically account for this via proxy indices such as F10.7, but with variations in solar flux forecasts leading to orbit differences of multiple kilometres over just a few days, prediction of these quantities may represent one of the largest limiting factors in drag modelling. This has fundamental implications both in the short term, in the day-to-day management of operational spacecraft, and in the mid-to-long term, in determining satellite orbital lifetime.
In this work, we apply state-of-the-art time series forecasting techniques and deep learning architectures for the prediction of the F10.7 proxy on the days-ahead timescales relevant to space operations. The novel N-BEATS architecture is employed, which has been shown to achieve winning accuracies in the M4 forecasting competition, and is compared to operationally available forecasts provided by the ESA Space Weather Service Network.

Key words: Deep Learning, Time Series Forecasting, Space Weather, Space Debris
"

A Deep Learning Approach to Space Weather Forecasting for Orbital Prediction

The abundant availability of modern satellite imagery (including optical images and Synthetic Aperture Radar (SAR) data) allows a continuous monitoring of processes on the Earth’s surface. However, training automatic systems to analyze such data requires large amounts of training samples. This is rarely the case for Earth Observation tasks. This talk discusses an example of self-supervised learning that allows to drastically reduce the amount of training data while increasing the achieved performance. The core idea is that learning the proxy task of transcoding SAR images into optical images forces an employed conditional generative adversarial network (GAN) to implicitly distinguish between different land surfaces. Such a network can then be used to build a classifier with significantly fewer free parameters that generalizes well even when trained on a very small amount of labeled data. We train such a GAN on aligned Sentinel-1 and Sentinel-2 image pairs and then show that a pre-trained classifier using these features learned from transcoding outperforms classifiers that are trained from scratch in particular when only a very limited amount of labeled pixels are available for training.

Downloads

Next from ECAI 2020

Data Fusion of Activity and CGM for Predicting Blood Glucose Levels

Similar lecture

Blood Glucose Level Prediction as TIme-Series Modeling using Sequence-to-Sequence Neural Networks

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from ECAI 2020

Data Fusion of Activity and CGM for Predicting Blood Glucose Levels

Similar lecture

Blood Glucose Level Prediction as TIme-Series Modeling using Sequence-to-Sequence Neural Networks

Downloads