ELECTROCARDIOGRAM COLLECTION, PATTERN RECOGNITION, AND CLASSIFICATION SYSTEM SUPPORTING A MOBILE CARDIOVASCULAR DISEASE DETECTION AID A Thesis in Computer Engineering Submitted in Partial Fulfillment of the Requirements for the.
Download ReportTranscript ELECTROCARDIOGRAM COLLECTION, PATTERN RECOGNITION, AND CLASSIFICATION SYSTEM SUPPORTING A MOBILE CARDIOVASCULAR DISEASE DETECTION AID A Thesis in Computer Engineering Submitted in Partial Fulfillment of the Requirements for the.
ELECTROCARDIOGRAM COLLECTION, PATTERN RECOGNITION, AND CLASSIFICATION SYSTEM SUPPORTING A MOBILE CARDIOVASCULAR DISEASE DETECTION AID A Thesis in Computer Engineering Submitted in Partial Fulfillment of the Requirements for the Degree of Master of Science By Patrick R. DaSilva COMMITTEE Dr. Paul Fortier (Advisor) Dr. Howard Michel Dr. Kristen Sethares Dr. Honggang Wang 2 AGENDA Introduction Background Problem Statement Research Questions Detection Classification Performance Testing Research Organizational Questions Development Implementation Live System Testing Conclusion Questions 3 INTRODUCTION 4 HEART DISEASE CDC Division of Vital Statistics Heart disease accounts for 26% of total deaths in 2007 Preliminary 2009 data showed that heart disease was still the main cause of death Patients don’t understand symptoms Leads to multiple hospitalizations Associate with other ailments Costs 34 billion dollars yearly Puts a financial strain on our healthcare system 5 CURRENT NURSING METHODS Nurse-led HF clinics Home care nursing Visit frequency varies Depends on patient physical activity ability Still require face-face Patients don’t remember or understand instructions Visitations not real-time do not always happen when they are needed 6 SUPPORTED SYSTEM Focus is to support a mobile patient-centric, selfmonitoring and self-intervention system with a clinical nursing tool to aid in collaborative patientclinician chronic disease management 7 MOBILE CARDIAC HEART MONITOR SENSOR SUITE FSM Ctrl Cardiac Wellness Engine Synthetic Sensor Derivations Tx DSP Temp A/D Conversion SPO2 Analog Conditioning EKG Memory Cardiac Event Monitor Rx Position / Motion 8 PROBLEM STATEMENT Focus on the research, development, design, testing and validation of the electrocardiogram sensor RESEARCH QUESTIONS Is it feasible to autonomously detect and classify EKG arrhythmias to assist in a medical situational assessment and health status feedback? Will the system be able to differentiate a normal rhythm from an abnormal one? Will the system be able to perform in or close to realtime? 9 RESEARCH 10 WHAT IS AN EKG? Recorded electrical activity that represents the heart’s conduction system 11 HOW IS AN EKG COLLECTED? Limb Leads Lead I, Lead II, Lead III Precordial Leads V1, V5 V2, V3, V4, V6 Augmented Limb Leads aVR, aVL, aVF 12 Lead 5 Lead 3 Lead 12 WHICH CLASSIFICATIONS? Normal EKG (LII, V1) Sinus Bradycardia (LII) Sinus Tachycardia (LII) Supraventricular Tachycardia (LII, V1) Slow-Fast Rate Atrial Fibrillation (LII) Irregular Rhythm Atrial Flutter (LII) First-Degree AV Block (LII) Type 1 Second-Degree AV Block (LII) Junctional Rhythm (LII, V1) Abnormal P Only Type 2 Second-Degree AV Block (LII) Third-Degree AV Block (LII) Bundle Branch Block (V1) Premature Ventricular Complex (LII, V1) Abnormal QRS Only Ventricular Tachycardia (V1) Ventricular Fibrillation (V1) Asystole (LII, V1) Abnormal P and QRS 13 HOW IS AN EKG PERFORMED TODAY? Stress tests are performed to monitor the effect of exercise on the heart Motion artifacts appear as various peaks or HF noise which affect detections Artifacts as well as an increased heart rate ultimately affect classifications Research has been done by others to eliminate such noise Software will assume patient is at rest 14 WHAT ARE THE PIPS? R P T Q PR Interval S Characteristic Description P wave 3mm high 0.12 seconds long upright QRS complex g.t. 5mm high 0.06 – 0.12s long upright T wave g.t. 0.5mm high upright U wave upright PR interval 0.12 – 0.2s long Atrial rate 60 – 100bpm Ventricular rate 60 – 100bpm 15 WHAT ARE CURRENT PIP EXTRACTION METHODS? Pan Tompkins method Developed in 1985 Non-linear transformation Jiapu Pan and Willis Tompkins Detection rule set further developed by Patrick Hamilton and EPLimited Finds QRS complex, but further filters required to find P and T waves Y(n) EKG Low Pass High Pass d dt 2 1 32 32 Z(n) 1 16 WHAT ARE CURRENT PIP EXTRACTION METHODS? (CONT.) Dyadic Quadratic Spline Wavelet Transform Q1 Q2 Q4 Q5 Uses 5 FIR filters running in parallel Q3 Focuses on time frequency analysis Separates EKG characteristics into various scales A peak’s delayed location is found when there is a zero crossing between two local maxima and minima peaks on a subset of the filter outputs Onsets and offsets calculated based on maxima and minima onsets and offsets 17 DEVELOPMENT 18 SYSTEM BLOCK DIAGRAM Filter 1 QRS Detect Filter 2 ADC Filter 3 Filter 4 Filter 5 Threshold Post-Detect Classify P/T Detect 19 FILTER DESIGN AND IMPLEMENTATION 𝐺 𝜔 𝑄𝑗 𝜔 = 𝐺 2𝜔 𝐻 𝜔 𝐺 2𝑗 −1 𝜔 𝐻 2𝑗 −2 𝜔 ⋅⋅⋅ 𝐻 𝜔 𝑗=1 𝑗=2 𝑗>2 20 FILTER DESIGN AND IMPLEMENTATION (CONT.) Constant 62 millisecond delay MIT-BIH Record 100 Filtered 21 DETECTION DESIGN AND IMPLEMENTATION QRS complex and P/T wave detection schemes run in parallel Decided by thresholds Shared Buffer Possible QRS peaks stay on the QRS side Possible P/T peaks may migrate to the QRS side QRS Detect Threshold Post-Detect P/T Detect 22 DETECTION THRESHOLDS Autonomous threshold technique used to find modulus peaks on filter outputs 4 sets of 2 2 sets for QRS (1,2) 2 sets for P/T (3,4) 𝑃𝑂𝑆𝑇𝐻𝑅𝑆1 = 𝑚𝑒𝑎𝑛 𝑃𝑂𝑆𝑃𝐸𝐴𝐾𝑆, 4 ∗ 𝑇𝐻𝑅𝑆𝐶𝑂𝐸𝐹1 𝑃𝑂𝑆𝑇𝐻𝑅𝑆2 = 𝑃𝑂𝑆𝑇𝐻𝑅𝑆1 ∗ 𝑇𝐻𝑅𝑆𝐶𝑂𝐸𝐹2 𝑃𝑂𝑆𝑇𝐻𝑅𝑆3 = 𝑃𝑂𝑆𝑇𝐻𝑅𝑆2 ∗ 𝑇𝐻𝑅𝑆𝐶𝑂𝐸𝐹3 𝑃𝑂𝑆𝑇𝐻𝑅𝑆4 = 𝑃𝑂𝑆𝑇𝐻𝑅𝑆3 ∗ 𝑇𝐻𝑅𝑆𝐶𝑂𝐸𝐹4 𝑁𝐸𝐺𝑇𝐻𝑅𝑆1 = 𝑚𝑒𝑎𝑛 𝑁𝐸𝐺𝑃𝐸𝐴𝐾𝑆, 4 ∗ 𝑇𝐻𝑅𝑆𝐶𝑂𝐸𝐹1 𝑁𝐸𝐺𝑇𝐻𝑅𝑆2 = 𝑁𝐸𝐺𝑇𝐻𝑅𝑆1 ∗ 𝑇𝐻𝑅𝑆𝐶𝑂𝐸𝐹2 𝑁𝐸𝐺𝑇𝐻𝑅𝑆3 = 𝑁𝐸𝐺𝑇𝐻𝑅𝑆2 ∗ 𝑇𝐻𝑅𝑆𝐶𝑂𝐸𝐹3 𝑁𝐸𝐺𝑇𝐻𝑅𝑆4 = 𝑁𝐸𝐺𝑇𝐻𝑅𝑆3 ∗ 𝑇𝐻𝑅𝑆𝐶𝑂𝐸𝐹4 23 QRS DETECTION Filters Q1, Q2, Q3 to find QRS Three simultaneous zero crossings correlate to a QRS peak Filter Q2 is used to locate the onset and offset of each QRS MIT-BIH Record 100 Filtered 24 CALCULATING PULSE RATE Ventricular heart rate calculated at this point as the mean of the last 8 heart rates found by getting the difference between successive R peaks MIT-BIH Record 100 25 P AND T WAVE DETECTION Filters Q3, Q4, Q5 to find P and T waves Found peak on Q3 && Q4 || Q4 && Q5 Detected as ‘Blip’ wave MIT-BIH Record 100 Filtered P,T, or U wave decided in postdetection Onset and offset are calculated similar to QRS onset and offset using Q4 26 POST DETECTION 27 DETECTION DELAYS QRS detection delay: 262ms to 462ms P/T detection delay: 162ms to 362ms 62 milliseconds Blanking Window Lag seen after 165 beats per minute Filter Delay Lag seen after 129 beats per minute 200 milliseconds for QRS, 100 milliseconds for P/T Future Values 200 milliseconds Finding second modulus peak Finding offset 28 CLASSIFICATION DESIGN AND IMPLEMENTATION 1. Examine the Rate 2. Examine the Rhythm 3. Examine the axis, intervals and segments 4. Examine everything else 29 CLASSIFICATION DESIGN AND IMPLEMENTATION (CONT.) 1. Examine the Rate 2. Examine the Rhythm 3. Examine the axis, intervals and segments 4. Examine everything else Classification NSR Sinus Bradycardia Sinus Tachycardia SVT Atrial Fibrillation Atrial Flutter 1st Degree AV Block 2nd Degree Block Type 1 (Wenckebach) Junctional Escape Rhythm 2nd Degree Block Type 2 3rd Degree Block (Complete Block) Bundle Branch Block PVC Ventricular Tachycardia Ventricular Fibrillation Asystole Abnormality/Rhythm Sinus Rhythm Atrial Arrhythmias/Sinus Rhythm Atrial Arrhythmias/Sinus Rhythm Atrial Arrhythmias/Supraventricular Rhythm Atrial Arrhythmias/Supraventricular Rhythm Atrial Arrhythmias/Supraventricular Rhythm Conduction Abnormalities Conduction Abnormalities Ventricular Arrhythmia/Supraventricular Rhythm Conduction Abnormalities Conduction Abnormalities Conduction Abnormalities Ventricular Arrhythmias/Ventricular Rhythm Ventricular Arrhythmias/Ventricular Rhythm Ventricular Arrhythmias/Ventricular Rhythm Ex. Sinus Rhythm is present when there is a 1:1 P wave ratio with all P waves being upright 30 CLASSIFICATION N-ARY TREE 31 CLASSIFICATION DELAY Overall classification delay is one beat Takes three normal beats for a normal classification MIT-BIH Record 16773 (NSR) 32 PERFORMANCE TESTING Test setup Test procedure Resample signals to 250Hz Start at 20 seconds into signal Stop at 10 minutes into signal Count hits and misses Test result type MIT-BIH Arrhythmia Database MIT-BIH Normal Sinus Rhythm Database Positive Abnormal Classification (PC) Positive Unknown Classification (PU) Positive Normal Classification (PN) Negative or missed Abnormal Classification (NC) Negative or missed Normal Classification (NU) Test Metrics Classification Hit Ratio (%) ((PC+PN)/(PC+PN+NC))*100 Normal Hit Ratio (PN/(PN+NU))*100 33 PERFORMANCE TESTING RESULTS Record 100 101 102 103 106 107 108 109 111 124 207 219 222 223 231 16265 16272 16773 PC 1 14 4 3 153 26 2 PU 5 9 378 PN 8 179 NC 2 15 22 36 NU 723 442 130 666 542 680 6 8 1 28 77 736 32 36 344 815 679 476 554 4 5 33 319 311 38 446 536 1 4 673 699 691 548 322 Class % 100 100 99.47 NA 28 100 0 99.03 99.85 94.44 90.18 0.54 13.51 62.11 48.27 NA NA NA Norm % 1.09 28.82 0.00 2.20 3.90 NA 0.00 NA NA 0.00 0.00 NA 0.00 0.00 NA 31.04 6.48 58.07 34 INVALID NORMAL CLASSIFICATIONS Classified Normal as Abnormal Sinus Rhythm Classified Normal as Sinus Tachycardia (100, 101, 103, 106, 223, 16265, 16272, 16773) Morphology calculations landed outside of textbook normal Still found a sinus rhythm, just wider/taller waves (16265, 16773) Software was correct after a further look into the signal Classified Normal as PVC (219, 223, 16265) Missed P waves, low ST segments that look like inverted T waves, same ASR issues Still underlying normal heart rate Classified Normal as Unknown High frequency noise that carried over to filters (108, 222) Missed P waves, low ST segments cause inverted T waves Tall Narrow P waves in close proximity to QRS (222) 35 ABNORMAL CLASSIFICATIONS PVCs classified as Unknown Incorrect QRS polarity detection (106,109) Missing P waves and underlying sinus rhythm (109,219) Really wide QRS complex (124) Triplet PVCs at rate less than 100 BPM (124) Atrial Fibrillation classified as Unknown Fibrillatory waves’ amplitude too low to detect atrial rate (219,222) Atrial Flutter classified as Unknown P waves tall, wide, close together cause incorrect P wave detections (222) 36 FIXING ASR To fix ASR issue, slightly modified textbook normal morphology MIT-BIH Record 103 before (top) & after (bottom) P wave height and width, PR interval Increased accuracy by an average of 38.06% More reliable fix Use a frequent Pattern Tree in a learning phase to learn the patient’s normal sinus rhythm Accelerometer input 37 INVALID ATRIAL FLUTTER Fix to Atrial Flutter classified as Unknown Create separate detection with same filter or separate filter/detection scheme all together MIT-BIH Record 222 38 P/T DETECTION INTERFERENCE MIT-BIH Record 222 To fix the Normal as Unknown classifications Add smoothing filters on front end Independent filter output buffers for QRS and P/T detection 39 CLASSIFICATION MODIFICATIONS 40 PERFORMANCE TESTING SUMMARY Detection modifications still need to be made Classification modifications were made Improved accuracy Auto-Threshold allowed for QRS detection after pacemaker stopped pacing to still detect QRS complexes Pacemaker spikes detected as QRS complexes Wavelet Transform still allows QRS complexes to be detected among HF baseline drifts 41 PROPOSED FUTURE MODIFICATIONS ADC Threshold12 Smooth Filter QRS Detect Filter 5x Post-Detect Threshold34 Classify P/T Detect 42 LIVE SYSTEM TESTING 43 LIVE SYSTEM TESTING Bytes Before After .text .data .bss Program Memory Usage 32596 336 30048 32932, 6.3% 32448 336 30048 32784, 6.3% Run Length (16 MHz Clock) Initialization: 100 us Classification Not Found: 150 – 210 us Classification Found: 460 – 550 us 44 LIVE SYSTEM TESTING DEMO 45 CONCLUSION 46 PROBLEM STATEMENT Focus on the research, development, design, testing and validation of the electrocardiogram sensor RESEARCH QUESTIONS Is it feasible to autonomously detect and classify EKG arrhythmias to assist in a medical situational assessment and health status feedback? Will the system be able to differentiate a normal rhythm from an abnormal one? Will the system be able to perform in or close to realtime? 47 RESULTS Proved feasibility to autonomously detect arrhythmias, differentiate between normal and abnormal rhythms all in real-time Contributed an additional combination of detection and classification software geared towards classifying normal sinus rhythm as well as 15 other abnormal classifications in realtime on an embedded platform with the ability to become mobile once integrated with other noise canceling sensors/filters 48 SUGGESTIONS FOR FUTURE WORK Further research and modifications needed for P and T wave detections Research required into a frequent pattern tree for classification with a learning phase 49 QUESTIONS 50