TEAMED (Technology Evaluating And Measuring Emotional Dysregulation)

Green Dot Introduction
Green Dot Approach
Green Dot System Organisation
Green Dot Sample Data
Green Dot Evaluation Results


Emotional dysregulation is a key feature in a variety of conditions. These include brain injury and the dementias, as well as severe and enduring mental health problems such as schizophrenia and bipolar disorder. Medical management predominates but is of limited use in neurological conditions such as acquired brain injury and the dementias. Increased survival rates and longevity mean that these conditions are an increasing public health concern, affecting 11% of the adult population. Treatment of emotional dysregulation must be founded on accurate measurement of emotional state. By increasing awareness of patient mental state, medical staff are enabled to be more supportive and patients can use feedback to better regulate their emotions.

The TEAMED project (Technology Evaluating And Measuring Emotional Dysregulation) developed, integrated and assessedg a variety of technologies to establish a patient’s emotional state. The desired outcome was for patients with emotional problems to receive better care, to achieve a better quality of life, and to self-manage their conditions better.

The project was funded by the Digital Health and Care Institute from November 2015 to August 2016, though further evaluation is ongoing. The collaborators were the University of Stirling (Ken Turner and Evan Magill), Rapport Network CIC (Gary Cornelius) and the Brain Injury Rehabilitation Trust (Brian O'Neill).


Physiological Monitoring:
Using results from the MATCH and PAM projects, physiological monitoring has been developed using standard components (accelerometers, activity sensors, galvanic skin response sensors, smartphones, etc.). Commercial technologies that have been integrated include a variety of smartwatches (e.g. Microsoft Band, Basis Peak) and low-power wireless devices in general. The aim was to use these devices in a novel and integrated manner to detect agitation, anxiety, mounting anger, distress, lassitude, poor sleep, etc.
Emotional Monitoring:
A multi-faceted approach has been followed to allow a more accurate assessment to be made of emotional state – particularly emotional distress and mood swings. Machine learning techniques have been adapted to allow emotional state to be predicted from sensor and clinical input.
Emotional Feedback:
The results of emotional monitoring are provided to patients in a simple and helpful form, allowing patients to self-manage their condition and to reduce the number and severity of emotional episodes. A more detailed analysis is provided securely to medical staff to help them identify trends, to anticipate significant deterioration, and to plan timely interventions.
Integration and Assessment:
The solutions to be developed by the project are generally applicable to a wide range of mental health problems. However to give the project concrete focus, the project results are being evaluated with brain-injured patients at the Brain Injury Rehabilitation Trust in Glasgow.

System Organisation

The overall structure of the TEAMED system is as follows:

Patient Database (InfluxDB):
Key data is securely captured in near-real time and stored in an InfluxDB time series database. Data is aggregated per minute and made available for further processing within ten minutes.
Smartwatches (Microsoft Band, Basis Peak):
Two kinds of smartwatch are currently supported: the Microsoft Band 2 and the Basis Peak. The focus is on collecting data of use in predicting emotional dysregulation: heart rate, heart rate variability (Band only), skin temperature, skin resistance and step count (as an indication of agitation). In addition, daily sleep quality and sleep total are stored.
Desktop Application (Teamed Admin):
This Java application runs on one or more desktop computers used by clinical staff. It is the primary interface for managing patients and their physiological monitoring. There is one module for each type of smartwatch used (Microsoft Band and Basis Peak). The data analyser module classifies physiological, sleep and report data in order to predict the risk of emotional outbursts. The chart display module provides a detailed visual breakdown of patient data. The alert display module is used to provide feedback to patients and staff about the risk of an emotional episode: either via a mobile phone or via a biofeedback device (the EasyBulb coloured light).
Phone Application (Teamed Patient):
This Android application runs on the patient's mobile phone. It is used to display the predicted risk of an emotional episode, thus providing feedback to patients so that they can improve their emotional state. The risk of an episode is shown as a coloured 'traffic light' from blue (low risk) to green (medium risk) to red (high risk) - the same colour scheme as used with the biofeedback light. When used with a Microsoft Band, the Teamed Patient application uses the separate Microsoft phone application to collect and store the raw physiological and sleep data. When used with a Basis Peak, the Teamed Patient application does not need to deal with raw data as this is automatically collected by the separate Basis phone application.
Data Analysis (Weka):
A configurable framework analyses the physiological and sleep data collected from the smartwatches. A further input to the analysis are the reports of aggressive behaviour collected by clinical staff. These records are used to train the machine learning algorithm, both as inputs (recent aggression may predict further outbursts) and as outputs (the target predictions of a classifier). Data mining techniques analyse features extracted from the data. These features are then used to train machine learning algorithms. Extracted features are analysed by the Weka data mining tool. This results in a classifier model that takes recent data and makes a prediction of future emotional dysregulation. The classification approach is flexible and is parameterised by multiple kinds of features and by selectable evaluation periods.

Sample Data

The following diagram shows the kind of data that is collected. This shows charts for a hypothetical patient with a 45% risk of an emotional episode:

Sample Data

Evaluation Results

A small-scale evaluation has been carried out in order to gain confidence in the approach and to establish whether emotional episodes can in practice be predicted. The evaluation was conducted with volunteer patients at BIRT (Brain Injury Rehabilitation Trust) in Glasgow. Further evaluation is ongoing in conjunction with BIRT.

Three patients at BIRT agreed to participate in the evaluation. The participants wore a Basis Peak or Microsoft Band watch and carried a mobile phone in a belt pouch. Data was collected over a period of 24 weeks. The total amount of data collected was 178 potential days of physical data (211,844 actual records) and 163 potential days of sleep data (105 actual records). In addition to this, medical staff entered 121 reports of aggressive episodes (423 actual records).

There were fairly frequent gaps in the data such that only 71% of potential physical data and 64% of potential sleep data was collected. For physical data the gaps ranged from tens of minutes to a few hours to a few days. Although this posed a challenge for the analysis it is probably typical of what can realistically be achieved in a clinical setting.

The overall aim of the work was to predict episodes of aggressive behaviour. A variety of classification approaches was investigated using different lengths of training data. The most accurate approach proved to be a Partial Decision Tree coupled with a cost matrix to compensate for the considerable imbalanace in data classes. (Periods with episodes are only a few percent of periods with normal behaviour.) The optimal cost for misclassifying an episode is computed automatically, though this can be fine-tuned manually.

The following table shows prediction performance averaged across all participants when using physical data. This gives the prediction accuracy for episodes and normal behaviour. The variation is shown for different training and actual periods: 1, 2 or 4 weeks for training and then actual performance in the subsequent 1, 2 or 4 weeks. The variation is also shown with different prediction periods: 1, 2, 3 or 4 hours ahead. Useful results were achieved with 4 weeks of training data and predictions 3 to 4 hours ahead. This gives medical staff adequate time to monitor a patient at risk and to intervene as appropriate. It suggests that the mechanisms that cause aggressive behaviour have a measurable physiological effect a few hours ahead of the actual aggression. Although performance is actually better with 2 weeks of training data and predictions up to 1 hour ahead, this would not give much warning of a potential problem and so was not selected.

  Training/Actual (weeks)
  1 2 3
Prediction (hours) Episode Normal Episode Normal Episode Normal
0-1 75% 16% 81% 42% 35% 64%
1-2 75% 16% 91% 9% 41% 53%
2-3 100% 17% 64% 64% 60% 45%
3-4 75% 16% 41% 59% 68% 42%
Prediction Accuracy with Full Validation of Physical Data

The following table shows prediction performance when using sleep data; note that the prediction periods here are in days. The best performance was for 2 weeks of training data and predictions 3 to 4 days ahead. It is perhaps surprising that reasonably accurate predictions can be made so far in advance. The reason is partly because several nights of poor sleep can trigger an episode, and partly because episodes a few days ago are a reasonable predictor of future episodes.

  Training/Actual (weeks)
  1 2 3
Prediction (hours) Episode Normal Episode Normal Episode Normal
0-1 25% 84% 82% 42% 47% 59%
1-2 60% 9% 82% 63% 90% 21%
2-3 25% 75% 80% 60% 72% 38%
3-4 60% 91% 82% 68% 69% 49%
Prediction Accuracy with Full Validation of Sleep Data

After reviewing prediction results from the evaluation, staff of the clinical partner (BIRT) judged that the accuracy was acceptable and that the approach would be useful. Indeed it is a big improvement on the current situation where episodes are virtually unpredictable and cannot easily be anticipated by the staff. Furthermore, BIRT were pleased that they would receive adequate warning of possible aggressive episodes. BIRT report that when patients become aroused and aggressive they respond well to being 'talked down' and encouraged to relax. There is therefore a real opportunity for intervention by medical staff to head off an episode. Further work is ongoing to consider actual interventions based on the automated predictions.

Up Arrow Up one level to Ken Turner - Research Projects

Web Ken Turner Home   Email    Search Search Web Pages

Last Update: 29th May 2017