heart attack prediction using wearables
TRANSCRIPT
Heart Attack Prediction
Heart Attack Prediction
-Ankita Singh and Harshitha Chidananda
● Heart attacks are the primary cause of increasing death-rate.
● Enormous amount of healthcare data is generated by various tracking devices (FitBit).
● Is there a way we could use it to predict health ailments (Heart Attacks)?
MOTIVATION
Aim● Leverage machine learning
algorithms to predict heart attacks.
● Propose a next generation wearable device to facilitate such a prediction.
Open Questions?● Is there a way to predict an ailment on the basis of the everyday activity?● Can the information provided by fitbit/other health trackers be used for
predicting a future health issue?● What extra features could be included to make a more informed analysis?
METHODOLOGY
● Use the data collected from various hospitals to create an ensemble model for heart attack prediction.
● Analyze the attributes of the data collected by FitBit.● Do a comparative study of the 2 datasets and propose additional attributes to be
included for the gen-next wearables.
DATASET
Dataset (Collected from hospitals)● UCI Machine Learning Repository-Heart Disease Dataset● Number of Records: 303● Number of features: 75● Features used: 13● Training: 250 Records● Testing: 53 Records
DATA PREPROCESSING
● Normalized the data to a standard normal distribution.● Used PCA as a feature extraction technique to select principal components .● Plotted the variance explained by the principal components to select the
necessary principal components.● 13 out of 75 features were selected to build the models.
PRINCIPAL FEATURES
● Age● Sex● Chest Pain Type● Resting blood pressure● Cholesterol● Fasting blood sugar(fbs)● Electrocardiographic results
● Maximum heart rate achieved(thalach)● Exercise induced angina● Depression induced by exercise● Slope (related to exercise behavior)● Number of vessels colored by fluoroscopy● Thal
MODELS● Decision Trees● SVM● Neural Networks● Ensemble (Bag Boosting)
BASELINE RESULTS
● Decision Trees - 69.8%● SVM- 75.47%● Neural Networks -
77.35%● Ensemble - 80.30%
FitBit Dataset● Used the fitbit public APIs to export self-generated physical activity.● A CSV file with a total of 17,568 observations.● Available Features
■ Heart Rate■ Number of steps walked■ Average elevation■ Calories burnt■ Sedentary minutes■ Body mass index■ Fat■ Sleeping Pattern
Missing Features
● Chest Pain Type● Resting blood pressure● Cholesterol● Fasting blood sugar(fbs)● Other specific cardiological Features
Feature Elimination
PROPOSED SOLUTION● Eliminating features which are
infeasible to be gathered using a wearable.
● Proposing new features to be added in FitBit
● Comparing the performance of the proposed wearable to the existing technology.
Proposed Features:
★ Blood Pressure(trestbps)★ Blood sugar(fbs)★ Resting electrocardiographic results★ Maximum heart rate achieved
Monitoring Blood Pressure● Until Sep 2014, FitBit provided an option to manually enter the monitored BP
readings into the health logs.● Existing wearables to monitor BP?
■ INDIEGOGO■ QUARDIOARM
● Integrate the data collected from these apps with the FitBit Logs.
INDIEGOGO
Monitoring Blood Sugar Level ● Basic members are given the ability to log three
glucose measurements per day.■ Morning■ afternoon ■ evening
● Premium Fitbit.com member (for $49.95 per year) you can log unlimited glucose measurements,
● But it does not gives the ability to make notes about the readings ("post meal" for example).
● Integrate with already existing Sugar monitoring wearables?
Resting Electrocardiographic Results + Max Heart Rate
QardioCore - A wearable which guarantees multi sensor ECG .
Prediction ResultsFeatures
● age● sex● trestbps● fbs● restecg● Max heart rate
Conclusion- We did an extensive analysis of the data collected from various hospitals to
predict the heart-attacks. (80.30%)- Analyzed the relevance of everyday data accumulated by FitBit.- Proposed a set of features to be supported by the existing hardware to facilitate
such health predictions.- Proposed model achieves an accuracy of 69.8%.
Future scope
- Other health issues- Improve accuracy- Design hardware- Integrating with existing
wearable
- Only 300 records.- Feasibility??- Reliability of the data
recorded by Fitbit.
Limitations
Strengths- Successful in predicting heart attack.- Synchronized with everyday activity patterns.- Easy and economical health monitorization.