1 / 36

Predicting clinical events using non-wearable sensors in elderly

Predicting clinical events using non-wearable sensors in elderly. Mihail Popescu MU Informatics Institute. Challenges in medical pattern recognition. Challenge no 1 : Hard to get data in sufficient quantity and quality Patient confidentiality (HIPAA) Hard to perform experiments

infinity
Download Presentation

Predicting clinical events using non-wearable sensors in elderly

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Predicting clinical events using non-wearable sensors in elderly Mihail Popescu MU Informatics Institute

  2. Challenges in medical pattern recognition • Challenge no 1: Hard to get data in sufficient quantity and quality • Patient confidentiality (HIPAA) • Hard to perform experiments • insufficient and incomplete data • Algorithm validation is difficult • Possible solution: hospital data warehouse

  3. Challenge no. 2 • Hard to obtain data for the “other” class  • severe class imbalance problem • hard to train a 2-class classifier, • Ex: • if we want to detect falls in elderly, we can’t collect fall data • If we want to detect heart attacks, we can’t provoke them  use methods that do not require training (expert systems, fuzzy rules) or one-class classifiers (anomaly detection)

  4. Challenge no 3 • Data = mixed numeric and symbolic (categorical) • Example: P1=(ICD9: 232.2, 421, age:62,chlesterol: 200, smoke:Y) P2=(ICD9: 230, 430, age:69,cholesterol: 120, smoke:N). Question: what is d(P1, P2)? use ontologies, cathegorical distances ( Burnaby, Goodal, etc) and relational algorithms (VAT, relational fuzzy c-means)

  5. Brief event ontology • Event • Clinical event • Chronic event • Abnormal blood pressure (BP) • Arthritis pain • Angina pain • Depression • Acute event • Fall • Medication adverse effect • Unspecified clinical event (“do not feel well”) • Non-clinical event (visitor, interesting book)

  6. Data source: TigerPlace • Location: Columbia, MO, USA • Mission: Aging in place • residents stay as active and functionally independent as possible • What is there: • First apartment instrumented 3 years ago. • Currently, 17 apartments on line • Sensors: • Present: motion, bed. • To come: video (silhouette sensor) and acoustic (fall detector)

  7. Fall Detection - using a two class classifier - using a fuzzy rule system

  8. Introduction • Fall detection approaches • Wearable devices (accelerometers, etc) • Non wearable: • Video sensors (cameras…) • Audio sensors (microphones…) • Others (radar, IR, magic carpet, etc) • The approaches are complementary: • Wearable sensors work outside (in the garden) • Non-wearable sensors are less intrusive and more suitable for seniors with mental disabilities • Audio sensors work during the night (bathroom visit)

  9. z Mic 1 Data Acquisition Card NI 9162 Fall signal (phone call, email) To caregiver 2 ft Mic 2 Motion detector 2 ft Mic 3 Microprocessor board x y FADE- Acoustic Fall Detection System Intended FADE architecture • Privacy concern: • FADE will be encapsulated with only (wireless) “fall” signals going out. • No sound will be stored. • Main technical problem: false alarms • Use an array of sensors for better location and confirmation • Use an integrated motion detector

  10. Available data • Falls performed by a stunt actor instructed to fall as an “elderly person” • Each fall session: • 10-15 min long • Had 3-5 falls • 6 fall sessions= 23 falls, 1.3 hours total • 1 extra session with 14 fall and 25 false alarms (steps, table knocks, object drops) was recorded for algorithm training • The training data was extracted manually in files with 1000 samples (1 s long)

  11. Methodology • 0. Consider windows with N=1000 samples and 0.5 overlap • 1. Signal preprocessing (for each channel) • Wiener filter • windows w with energy Ew< ETHR “no fall”

  12. Methodology (cont.) • 2. Remove false alarms using height • Perform spectrum cross-correlation • Compute the delay between the channels • Label the window “no fall” if 12>0 (signal came from above 2ft) • 3. Extract the cepstral features (mfcc) with C=7 (C0 was not used)  6 features • 4. Identify the sound using the NN. • 5. A “fall” has to be identified in both channels

  13. Results • Noisy environment: the nurse (standing, not shown) was instructing the actor

  14. Results for the NN (cont.) • ROC was obtained by varying ETHR • The false alarms were reported vs. time and not versus total number of false alarms (unknown) • Best performance: 100% detection with 5 alarms/hour  too much! an acceptable rate could be 1 false alarm/day (two order of magnitude lower)  How do we get there?

  15. More intuitive sound features Fall Bag drop Door knock

  16. Use sub-band energy ratios (ERSB) features • ERSB1(0-330Hz), ERSB2(331-2205Hz), ERSB3(2206-5513Hz)

  17. Fuzzy rule system (FRS) • If ERSB1=HIGH1and ERSB2 = LOW2and ERSB3=LOW3then “fall” • If ERSB1=HIGH1and ERSB2 = HIGH2and ERSB3=LOW3then “no fall” • … • If ERSB1=LOW1and ERSB2 = HIGH2and ERSB3=HIGH3then “no fall”

  18. FRS results • Dataset: 30 falls+50 fa • FRS performed as well as cepstral features+ nearest neighbor

  19. Abnormal blood pressure prediction

  20. A typical apartment sensor network • 5-8 motion sensors • 1 bed sensor (motion in bed-restlessness, pulse, breathing) • Other sensors: • Stove (temperature) • Refrigerator • Kitchen cabinets • Drawers • Video and audio sensors are under development

  21. The Data Logger • The sensors transmit events (on, activated) wirelessly to the data logger that adds time stamps and stores the events in a database • The sensor with continuous values (pulse) are quantized in 4 levels (we use only level 1 here) • Ex: level 1: move 5 seconds, level 2, move 10 seconds, etc… • Typical database record (firing):

  22. Question • Is it possible to correlate the sensor reading with abnormal clinical events? • Why?: alert nursing staff to check the resident (elderly do not report their status…) • Intuition: If the patient does not feel well he does not sleep well (during the night) and does not move as much (during the day) • This translates in: high restlessness during the night and low motion during the day

  23. Why pulse pressure (PP)? • PP=systolic BP – diastolic BP (mmHg) • PP is elevated (abnormal) when • Systolic BP is high • Diastolic BP is low (more often in elderly) • PP > 60 is associated with myocardial infarctions, renal and cerebral incidents • Problems • The threshold (60 mmHg)-normal PP, is questioned • The normal PP increases with age (ignored here) • It seems that mean arterial pressure might have been better • MAP~DP+PP/3

  24. Feature description • Divided (arbitrarily) the day in two • Night (9pm, previous day -7am) • Day (7am -9pm) • A better way would be to compute the go_to_bed and wake_up events (sleep duration !) • Used 4 features to describe the day of a resident: • Total night motion firings • Total day motion firings • Total day bed restlessness (level 1) • Total night bed restlessness (level 1)

  25. Available data • The study was retrospective not many BP readings were available Future solution: use a vital sign meter (Honeywell) • Out of the room: the resident was out of the room for more than 3 hours in the previous day the data was not used  Future solution: use firing density instead of the sum • Not that bad: there are hundreds of publications about classification of microarray data with less samples than this!

  26. Classifiers used • Divided data in two classes: abnormal PP (PP>=60) and normal (PP<60) • Used a classifier to predict the PP based on the previous day sensor readings: • Neural network M-M-1 • M=# of features (4 or 8) • Output: the degree of abnormality • Robust logistic regression: PP=f(feat1, …,featM) • Support Vector Machine (SVM) • Validation: • ROC curves and • leave-one-out cross-validation

  27. Results: classifier comparison Male1 Female1 • The robust regression seems to perform best in our conditions (insufficient data) • The NN did not have enough training data • We did not compute the ROC for the SVM

  28. Pulse pressure prediction

  29. Abnormal behavior pattern detection (“bad night”)

  30. One-class Classification Methods • Aka “Abnormality detection”, “novelty detection”, etc • Used where the “other” class is not available such as in intrusion detection, credit card fraud, medical surveillance. • Density methods: • estimate the density of the training data and set a threshold on this density • Trick: use a rejected fraction in training to remove possible outliers • Ex: Parzen density estimator, Gaussian model

  31. One-class Classification Methods - cont • Boundary methods: • Focus only on the boundary of the data • Deal better with small datasets • Ex: NN (nearest neighbor) and the SVDD (Support Vector Data Description)- an type SVM method • R= the distance from the center of the data set, “a”, to any support vectors • z is an outlier if ||z-a||>R

  32. Some experiments on Male1 data • What is an abnormal event: • Bad night complaints in the journal (only 2) • Bad nurse assessment during a day visit (as interpreted by Elena) • Abnormal pulse pressure • Sensor data was recorded hourly (bed restlessness, motion, heart rate, breathing) since we want to act as fast as possible (not next day) about 10,000 data points in total • Unfortunately: the abnormal events were recorded for the whole day • We focused on the nights, only want to detect a bad night.

  33. Comparison of SVM and SVDD • For data between 1am-2am. • The training data was smoothed with a running average over a month • SVDD (reject fraction 5%) does better than SVM : • Only 9 bad training cases • There might be more unlabeled “bad nights”

  34. Conclusions • Sound sensor arrays seem to be a viable technology for fall detection • Fuzzy rule systems work well for fall sound classification • One-class classifiers are a promising approach to medical surveillance

  35. Acknowledgement • Elena Florea • Yun Li • Eldertech team

  36. Thank you! • Questions?

More Related