Heart Helper ❤️

Made by Amin B. Z.

Note: Only longBeachVA.csv is used for training this model

This project is an attempt at making a machine-learning model that can predict if a person has heart disease without visiting a doctor with the following inputs:

• age (integer)
• sex (1 = male; 0 = female)
• chest pain location (1 = substernal; 0 = otherwise)
• pain is provoked by exertion? (1 = provoked by exertion; 0 = otherwise)
• relieved after rest? (1 = relieved after rest; 0 = otherwise)
• resting blood pressure (in mm Hg on admission to the hospital) (systolic) (e.g., if your blood pressure is 120/80, then enter 120)
• do you smoke? (1 = yes; 0 = no)
• has there been anyone in your family with heart disease? (1 = yes; 0 = no)
• maximum heart rate achieved

Research

The main issue is that some heart diagnosis tests are expensive and some of the cheap tests are not very accurate. EKGs (electrocardiogram) cost approximately $50, exercise stress tests cost $175+.⁴ Also, imaging tests cost between $500 and $2,000.⁵

I wanted to create a machine learning model that could accurately predict if one has heart disease via, what I call, “absolute” inputs (inputs that have no uncertainty behind them).

While searching the University of California, Irvine Machine Learning Repository, I found the Heart Disease Data Set that is used in training and testing the model.²

Approximately 610,000 people die of some type of heart disease every year in the US, and 47% of sudden cardiac deaths do not even occur in the hospital.⁶ The large proportion of sudden deaths occurring outside of hospitals means that people are not aware of their heart condition. People probably do not know about their heart condition because the cost of testing is too high and other low-cost tests could be misleading.

About the Data and Model

The data that I collected from the UC Irvine Machine Learning Repository¹ was somewhat incomplete. In an attempt to fix this issue, I used an imputer to complete the data by inferring from the already collected data, which is sometimes better than outright throwing the incomplete data away.

Also, the data had initially classified the types of heart disease on a scale of 0-4². 0 means heart disease is absent in the person. 1-4 means that there is some form of heart disease present in the person. I replaced values 1-4 with just 1 because my goal was not to distinguish between the different forms. I only wanted to predict if people had heart disease.

Sci-kit learn was the Python library that was used to create this model.³

The model's core is comprised of a pipeline that includes Kernel Principal component analysis (KPCA) followed by a Multilayer Perceptron Classifier (MLP).

The KPCA is used as a transformer in the pipeline. It essentially helps cluster, or transform, the different components, which are the different questions asked, of the training data in a nonlinear fashion. This is useful because otherwise, the data would be scattered in a nonmeaningful way, which would lower the accuracy of the model's predictions.

The MLP is used as a final estimator in the pipeline. It applies different weights to the different questions (let's refer to the questions as neurons), which means that some questions may be relied on more heavily as an indicator for heart disease. The MLP adjusts these weights based on the training data. One row of data from the training data is inputted to the MLP at a time, and the MLP compares its output with the expected output and calculates an error value, which is called stochastic gradient descent. This error value is then sent back, one neuron at a time, so that each neuron adjusts its weights as needed to minimize error, which is called backpropagation. The MLP repeats this learning process with each row of training data.

The pipeline uses the KPCA and MLP, in that order, to maximize the accuracy of the machine-learning model's predictions.

Results from 10 consecutive tests

Note: Tables are from same tests, they are just put in different tables. Also, there was no 'random-state' value set for these tests.

Data Recorded On: April 7, 2018 (Outdated!)

Predicting 0 (no heart disease):

Test ID	Precision	Recall	F1-Score	Support
1	0.79	0.71	0.75	21
2	0.68	0.94	0.79	16
3	0.77	0.77	0.77	22
4	0.83	0.86	0.84	22
5	0.78	0.82	0.80	22
6	0.83	0.75	0.79	20
7	0.81	0.91	0.86	23
8	0.77	0.81	0.79	21
9	0.86	0.90	0.88	20
10	0.89	0.76	0.82	21

Predicting 1 (heart disease):

Test ID	Precision	Recall	F1-Score	Support
1	0.65	0.73	0.69	15
2	0.93	0.67	0.78	21
3	0.71	0.71	0.71	17
4	0.73	0.67	0.70	12
5	0.69	0.64	0.67	14
6	0.71	0.80	0.75	15
7	0.78	0.58	0.67	12
8	0.71	0.67	0.69	15
9	0.85	0.79	0.81	14
10	0.75	0.88	0.81	17

Link to table for averages from both tables

Requirements to Run Program

• Install Python (version 2.7.14)
• Install packages: numpy (version >= 1.14.2), pandas (version >= 0.21), scikit-learn (version >= 0.19.1), imbalanced-learn (version >= 0.3.3)

References

1. Dua, D. and Graff, C. (2017). UCI Machine Learning Repository [http://archive.ics.uci.edu/ml]. Irvine, CA: University of California, School of Information and Computer Science.

2. Janosi, A., Steinbrunn, W., Pfisterer, M., & Detrano, R. (1989). Heart Disease Data Set [http://archive.ics.uci.edu/ml/datasets/Heart+Disease]. Budapest, Hungary: Hungarian Institute of Cardiology, Zurich, Switzerland: University Hospital of Zurich, Basel, Switzerland: University Hospital of Basel, Long Beach and Cleveland Clinic Foundation: V.A. Medical Center.

3. Scikit-learn: Machine Learning in Python, Pedregosa et al., JMLR 12, pp. 2825-2830, 2011.

4. http://www.choosingwisely.org/patient-resources/ekgs-and-exercise-stress-tests/

5. http://www.choosingwisely.org/patient-resources/imaging-tests-for-heart-disease/

6. https://www.cdc.gov/heartdisease/facts.htm