professional interfaces for communication and education purposes ... Communication Systems, June 2007. [7] R.Paradiso ..... [4] Haykin, Simon. Prentice Hall ...
MobiHealth 2010 International ICST Conference on Wireless Mobile Communication and Healthcare
An Intelligent System for Classification of Patients Suffering from Chronic Diseases Christos Bellos1, Athanasios Papadopoulos1, Dimitrios I. Fotiadis2, and Roberto Rosso3 1Foundation
for Research and Technology - Hellas, Biomedical Research, Ioannina Greece 2Dept. of Materials Science and Engineering, University of Ioannina, Greece 3TESAN Telematic & Biomedical Services S.p.A., Vicenza Italy
Contents • Necessity and Outcome of the System • State of the Art • Introduction to CHRONIOUS System • Main System Components – Pre-processing and Feature Extraction – Patient Profiler – Mental Support Tool – Decision Support System – Severity Estimation and Alerting
• Conclusions • Future Work
Necessity of the System • Need to perform all medical procedures and monitor chronically ill people remotely in the most efficient way according to the most updated international guidelines (in our case for COPD and kidney diseased patients) • Clinicians need to: – check the patient profile related information (history , latest measurements, prescriptions, appointments) – prescribe a patient plan: diet, drugs, exercises, lab tests and questionnaires) – set the patients’ alarms – query to the most updated medical knowledge
Outcome of the System Patients
• Lifestyle improvement • Active participation in monitoring and decision-making process • Helps patients to control their mental status
Healthcare Providers
• Improvement of the prognosis for the chronic diseases • Selection of the most appropriate treatment planning • New guidelines for patient remote monitoring
Healthcare Institutes
• Reduction of the number and duration of hospitalization • Increase of the home care assistance
State of the Art 1/2 • Several Chronic Disease Management tools are commercially available or supported by various healthcare services – Mainly home care systems – Very few of them provide wearable and mobile solutions for outdoors monitoring and disease management
• Very few of them, integrate decision support tools and alerting mechanisms for caregivers • Only a few of them are accompanied with patient and health professional interfaces for communication and education purposes
State of the Art 2/2 Systems utilizing similar technology based on the analysis vital and motion signals
Research projects utilizing techniques based on the development of Wireless Personal Area Network (WPAN)
MIThril (MIT Media Lab)1
Mobihealth5
CodeBlue (Harvard University)2
SAPHIRE6
Medical Embedded Device for Individualized Care (MEDIC)3
MyHeart7
Lifeguard system4
HEARTS8 U-R-SAFE9 Mobi-Dev10
[1] S. Pentland, IEEE Comp 37(5), 34-41, 2004 [2] K.Lorinz et al, IEEE Perv Comp 3(4) 16-23, 2004 [3] H. Winston et al, Artif Intel Med, 42, 137-152, 2008 [4] K. Montgomery et al, California EMBC 04, 240, 2004
[5] A. Van Halteren et al, Studies in Health Technology and Informatics, N.108, pp 181-193 IOS Press, 2004 [6] O. Nee, et al, IEE Proc CornSpecial issue on Telemedicine and e-Health Communication Systems, June 2007 [7] R.Paradiso, 28th IEEE EMBS Annual International Conference New York City, USA, 2006 [8] O. Faschimbecher, J. Epidemiol Comm Health, 56, 2002 [9] P. Rumeau et al, Int Assoc Geront Con Rio, June, 2005 [10] Mobi-Dev, 1st Inter Symp on Mobile Devices for Healthcare, Frascati, Italy, May, 2003
External Devices Spirometer Home Patient Monitor
Administrator
Glucometer Clinician HPM DB Central DB Blood Pressure
Internet Laboratory PDA DB
PDA
Weight Scale
Wearable
Env. Monitor
• Consists of a tight-fitting and washable shirt which provides the support for the stabilization of the sensor network • A data handler device is placed at the lower part of the shirt – accurate collection and management of the wireless (BT) transmission of the signals
• Categories of Sensors – – – – – –
Motion system – 3axis accelerometer ECG sensor (3 channels) Respiratory sensor (2 respiratory bands) Pulse Oximeter Temperature (Body and Ambient), Humidity Audio Sensor (Microphone)
Accelerometer Body Temperature sensor Ambient Temperature & Humidity Sensor
Pulse Oximeter sensor ECG electrodes Respiration bend (T)
Audio Sensor (Microphone)
Respiration band (A)
• Placed at Smart Assistant Device Consists of:
Aims to:
Feature Extraction
Identify and classify abnormal health events
Heterogeneous Data Fusion
Detect and evaluate a stressful situation
Decision Support System (DSS)
Provide alerts and/or advices to end-users
Mental Support Tool
Transmit data to Central system for further analysis
Severity Estimation and Alerting
Other Data Sources
Wearable Sensors
Pre-processing and Feature extraction
Mental Support Tool
Decision Support System (DSS)
Alerting Mechanism
Data transmission to Central system for further analysis
Alarms & Displays
Data Fusion
Pre-processing and Feature extraction
Sensor Data / Data Handler
Patient Inputted data (food, drug, activity)
Data from External Devices
Heterogeneous Data fusion Graphical User Interface
Patient Inputted Data Analyzer
Home Patient Monitor
Patient Profiler Output
Pre-processing and Feature Extraction • 2D signals collected from jacket’s sensors Electrocardiogram (ECG)
Respiratory Signals
Acceleration Signals
Pre-processing and Feature Extraction - ECG Preprocessing steps
Electrocardiogram Respiration Accelerometer
Related Activity
Removal of low frequency base line wandering (BW)
Linear filtering and Polynomial fitting techniques are employed
Removal of high frequency noise
The Daubechies (DB4) wavelet is employed
QRS Detection
1. Filtering the ECG signal with continuous (CWT) and fast wavelet transforms (FWT)1 2. Papaloukas et al. [2]
[1] YC Chesnokov, DN. Computers in Cardiology, 2006. [2] C. Papaloukas, D.I. Fotiadis, A. Likas, and L.K. Michalis, Journal of Electrocardiology, 2002.
Pre-processing and Feature Extraction – ECG Features Time domain features1 SDNN (msec): Standard deviation of all normal RR intervals in the entire ECG recording
SDANN (msec): Standard deviation of the mean of the normal RR intervals for each 5 minutes period of the ECG recording SDNNIDX (msec): Mean of the standard deviations of all normal RR intervals for all 5 minutes segments of the ECG recording pNN50 (%): Percent of differences between adjacent normal RR intervals that are greater than 50 msec r-MSSD (msec): Square root of the mean of the sum of the squares of differences between adjacent normal RR intervals over the entire ECG recording
[1] M. Bolanos, H. Nazeran, E. Haltiwanger, Engineering in Medicine and Biology Society, 2006.
Electrocardiogram Respiration Accelerometer
Pre-processing and Feature Extraction – ECG Features
Electrocardiogram Respiration Accelerometer
Frequency domain features The Low – Frequency band (LF) which includes frequencies in the area [0.03 – 0.15] Hz The High – Frequency band (HF) which includes frequencies in the area [0.15 – 0.40] Hz
Pre-processing and Feature Extraction – Respiration
Electrocardiogram Respiration Accelerometer
Respiration features Respiration Rate : The number of breaths per minute Tidal Volume (VT) : The normal volume of the air inhaled after an exhalation Vital capacity (VC) : The volume of a full expiration. This metric depends on the size of the lungs, elasticity, integrity of the airways and other parameters, therefore it is highly variable between subjects Residual volume (VR) : The volume that remains in the lungs following maximum exhalation
Pre-processing and Feature Extraction – Accelerometer
Electrocardiogram Respiration Accelerometer
Features
Methodology
Detect the postures of the patient
Posture detection is done by summing up 100 samples and analyzing them to find the highest value regarding the axes x, y or z
Count the number of steps in order to estimate the calorie burning rate and the effect of the exercise program
To recognize steps, peaks over a certain threshold are searched
Detect falling (tumbling), in order to trace the health status
To detect falls the difference between two consequent samples is measured and in case they exceed a limit, the event is characterized as ‘fall’
Feature Extraction Testing Environment
Data Collection Protocol Step 1: Healthy volunteers were recruited Step 2: Calibration procedure of the sensors (5 min)
Step 3: Validation against gold standard devices
Step 4: Recording of the collected signals at the several postures
• Standing • Lying • Falling down • Lay down slowly • Walking
Step 5 Annotation of the Dataset performed by experts
Data Collection • The test activities of the first phase will be done in Italy involving the main patients category which will take advantage of the highest benefits foreseen in terms of management of scientific and technical result Careggi Hospital Application field: COPD Patients involved: 40 patients ULSS 2 - Feltre Application field: KD Patients involved: 10 patients at hospital, 20 patients at home
Patient Profiler - Scope • This component enhances the personalization of the Clinical Decision Support System (CDSS) • Aims at identifying representative profiles of patients by clustering technique • Is a disease-specific web service which forms a vector containing the centroids of the cluster that the patient belongs to • The output vector enters the CDSS with an enhanced weight affecting the final decision.
Central DB PDA DB
Vector of the Profile Heart Rate
68
Steps
250
…
…
Activity Calories
122
Heterogeneous Data fusion
Decision Support System (DSS)
Some Indicative Attributes Several attributes have been identified as significant for characterizing the patients Attribute
Data Type
Attribute
Data Type
Day Period
String
Steps
Number (steps)
Heart Rate (HR)
Number (HR)
Falls
Number (events)
SpO2
Number (SpO2)
Systolic BP
Number
Humidity
Percent (%Humidity)
Diastolic BP
Number
Ambient Temperature
Number (oC)
Drug Condition
String
Body Temperature
Number (oC)
Total Calories
Number
Respiration Rate
Number
Total Lipids
Number (fat intake)
Lying
Number (minutes)
Activity Calories
Number
Standing
Number (minutes)
Activity Comments
String
Patient Profiler – Implementation (1/2) • Several clustering algorithms have been developed and tested for their efficiency • Clinicians have checked and assessed the output vector of each cluster for each clustering method • The output of the k-Means algorithm with five prespecified clusters has been chosen by the clinicians as the appropriate for the specific attributes and the testing dataset
Patient Profiler – Implementation (2/2) Algorithm EM Clustering Algorithm1 K-Means2 K-Means2
Hierarchical Clusterer3
Number of Clusters 4
2 (Prespecified) 5 (Prespecified)
5 (Prespecified)
[1] N. Mustapha, et al, European Journal of Scientific Research, 2009. [2] C.M. Bishop. Oxford, England: Oxford University Press, 1995. [3] Romesburg, H. Clarles, Cluster Analysis for Researchers, Pub. Co., 2004.
Clustered Instances 0
18
1
3
2
5
3
7
0
30
1
3
0
13
1
2
2
2
3
11
4
5
0
6
1
4
3
1
4
1
Mental Support Tool - Objectives • Detect a stressful situation by combining various parameters stored in PDA Database • Evaluate the identified situation by estimating a Stress Index and provide a respective advice to the patient • Evaluating the patient’s adherence to the system’s scheduled activities and recommendations • Keep patient’s moral high and motivating him/her to assist the disease management system • Provide tips once a day with short messages of clinical knowledge
Mental Support Tool – Block Diagram Patient Inputted data (food, drug, activity)
Data Fusion
Mental Support Tool
Sensor Data / Data Handler
Stress Evaluator
Alarm Viewer (GUI)
Implementation Steps
Attributes processing and Data Fusion
Constructing the Bayesian Network
Learning the Bayesian Network
Implementation
Learning will be achieved through: • Available data • Expert’s knowledge • Combination of both
Validation
Attributes processing and Data Fusion Specification of the attributes
• The attributes that affect the stress index and their weights have been specified based on the clinicians’ feedback • The specified parameters have been grouped to causes and symptoms of stress • Further processing of some attributes has been performed and meta-features have been extracted
Indication of the independent probability of causing stress Calculation of the transition matrix
• Further relations between causes and symptoms have been specified linking the attributes and combining their probability of stress presence
Indicative Attributes and Probabilities Attribute
Attribute Categorization
Smoking Environmental Noise and rowded/Noisy places
Causes Hypoglycaemia
Symptoms
Heart Rate Skin Temperature Breathing asynchrony Sleep Disturbances (Questionnaires) Mood (Questionnaires)
Activity Comments
Different States
Probability of causing Stress (%)
YES
90
NO
10
High
70
Medium Low YES NO High Normal Cool
25 5 85 15 85 15 35
Sweat YES NO YES NO Better Same Worst Feel sick, nausea Exhaustion, fatigue Discomfort in the chest, upper body, or jaw Irregular or extremely rapid heart beats None
65 90 10 62.5 37.5 5 25 70 18 23 23 28 8
Constructing the Bayesian Network • The Bayesian network has been designed.
WP 8
Learning the Bayesian Network • To this point both structure and parameters are defined based on expert’s knowledge • However is difficult for experts to quantify knowledge • It is important to update the model using real data1 – Data are also required for validation of the model – The more the available data, the better the quality of the resulting model
• Missing data can occur due to many reasons: – Sensor’s failure – Expert’s uncertainty on annotation
• There is a variety of methods for learning from missing data and have been applied (Antal et al. [2], Fengzhan et al. [3], Liao et al. [4])
[1] W. Liao and Q. Ji, Pattern Recognition, 2009. [2] P. Antal, et al., Artificial Intelligence in Medicine, 2004. [3] T. Fengzhan, et al, ICDM 2003. Third IEEE International Conference on Data Mining, 2003. [4] W. Liao and Q. Ji, Pattern Recognition, 2009.
Implementation • The specified parameters have been grouped to causes and symptoms of stress • The probability of causing stress by the presence or the absence of a specific state of each parameter has been specified. • We have implemented the Bayesian networks engine for inference and parameter learning: – – – –
The engine is based on the Junction Tree algorithm1. Implemented in C++. A C# wrapper is used to invoke the engine from the User Interface (UI). Developed for desktop and mobile applications and tested on Windows XP and Windows Mobile platforms.
[1] C. Huang and A. Darwiche, International Journal of Approximate Reasoning, 1996.
Mental Tool – Testing Environment
Validation Protocol Patients’ actions: Step 1: Patients wear the jacket during the setup phase (1 week collection of data) Step 2: Through validated questionnaires report stressful situations during the day Step 3: Classify the severity of the stressful episodes judging by experience
Clinicians’ actions: Step 1: Clinicians monitor the patient during the initial collection of data Step 2: Identify possible stressful episodes through clinical data and sensors’ values examination Step 3: Classify the severity of the stressful episodes
Decision Support System – Technical Roadmap Contains two classifiers running in parallel: – Expert system: The decision gives an explanation of the severity assessment • Rule based1
– Supervised Classification System: Detects and evaluates the episodes • Support Vector Machines2 • Random Forest3 • Multi-Layer Perceptron4
• Decision Tree5 • Naïve Bayes6 • PARtial decision Trees7
[1] A. Giacometti, E.K. Miyaneh, P. Marcel, A. Soulet, A Generic Framework for Rule-Based Classication [2] Cristianini N, Shawe-Taylor J. USA: Cambridge University Press; 2000 [3] L. Brieman. Machine Learning Journal, 45, pp. 5-32,2001 [4] Haykin, Simon. Prentice Hall, 1998 [5] J48, R. Quinlan, CA: Morgan Kauffman, 1993 [6] N. Friedman, D. Geiger, M. Goldszmidt, Machine Learning, 1997 [7] Eibe Frank , Ian H. Witten, Proceedings of the Fifteenth International Conference on Machine Learning, 1998
Severity Estimation and Alerting Level 1
Severity Estimation
Severity Level
No Action
Level 2
Level 3, 4, 5
DSS
Pop-up on PDA (simple advice that depends on patients profile)
Call Patient
Patient
Alarm Manager SMS Transfer Method
Contact User
E-Mail
Clinician GUI Clinician
Healthcare Assistant
24-h standby personnel
Conclusions • The project results to the development of an advanced multi-parametric intelligent system that fuses information effectively from various sources using intelligent techniques • Several features have been successfully extracted from Accelerometer, ECG and Respiration signals – In addition several data are being added through Graphical User Interfaces regarding food intake, drug intake and activity information
• The Decision Support System exploits both a neural network and a rulebased system running in parallel • The CHRONIOUS system achieves personalization of the severity estimation: – several profiles are created, describing with accuracy the health status of the patient, and used as an input to the Intelligence of the System
Future Work • Healthy volunteers and patients suffering from Chronic Obstructive Pulmonary Disease (COPD) and Chronic Kidney Disease (CKD) will be recruited to collect data • The weights of the various fused attributes will be validated and re-calculated • The developed classifiers and the respective algorithms will be re-trained in order to increase the accuracy of the system • The Patient Profiler component will be re-trained in order to conclude to efficient, with clinical meaning and disease-specific profiles