A Personal Health Agent for Decision Support in Arrhythmia Diagnosis

Abstract

We propose an architecture for a personal health agent (PHA) that combines machine learning and a Bayesian network (BN) for detecting and diagnosing heart disease, specifically arrhythmia. Machine learning (ML) is used for classifying a patient’s ECG signal. Four ML models, i.e. gradient boosting, random forest, multilayer perceptron and support vector machine, are compared and evaluated using a dataset of 5,340 records containing 12-lead ECG signals created from the Chapman-Shaoxing database. Among the four models, the gradient boosting model produces the best accuracy of 82.88% when classifying an ECG signal as either atrial fibrillation, other arrhythmia, or no arrhythmia. The detected pattern is integrated into a BN that captures expert knowledge about the causes of arrhythmia. The BN structure and parameters are informed by expert knowledge from the literature and evaluated using Pitchforth and Mengersen’s framework. The agent uses a decision support module to guide the diagnosis process. It suggests what questions to ask to increase certainty of the presence of arrhythmia, and it suggests what arrhythmia causes to follow up. This is achieved using sensitivity analysis and diagnostic Bayesian reasoning respectively. The architecture is evaluated using application use cases.

Appendix

Bayes’ Rule

$$\begin{aligned} Pr(A|B) = \frac{Pr(B|A)Pr(A)}{Pr(B)} \end{aligned}$$

(1)

where Pr(A|B) is the posterior probability of A given B; Pr(B|A) is the posterior probability of B given A; and Pr(A) and Pr(B) are prior probabilities of A and B respectively.

Entropy and Mutual Information

$$\begin{aligned} ENT(X) = - \sum P(x)logP(x) \end{aligned}$$

(2)

$$\begin{aligned} MI(X|Y) = ENT(X) - ENT(X|Y) \end{aligned}$$

(3)

Hyperparameter Selection Table

Algorithm	Hyperparameter	Options	Selected Options
Gradient Boosting	n_estimators	300, 500, 800	800
	criterion	friedman_mse, squared_error, mse	friedman_mse
	loss	log_loss, exponential	log_loss
	max_depth	1, 3, 10	3
Random Forest	n_estimators	300, 500, 800	800
	criterion	gini, entropy, log_loss	entropy
	max_depth	50, 100, None	None
	max_features	sqrt, log2, None	sqrt
SVM	C	0.5, 1, 1.5	1.5
	kernel	poly, rbf, sigmoid	rbf
	gamma	scale, auto	auto
	decision_function_shape	ovo, ovr	ovo
MLP	hidden_layer_sizes	N/A	(158, 100, 50)
	activation	identity, logistic, tanh, relu	tanh
	batch_size	auto, 64, 100	auto
	solver	lbfgs, sgd, adam	adam
	learning_rate	constant, invscaling, adaptive	adaptive
	max_iter	200, 500, 1000, 2000	500

ML Metrics: Accuracy, Precision, Recall, F1-Score

$$\begin{aligned} Accuracy = \frac{TP+TN}{TP+TN+FP+FN} \end{aligned}$$

(4)

$$\begin{aligned} Precision = \frac{TP}{TP+FP} \end{aligned}$$

(5)

$$\begin{aligned} Recall = \frac{TP}{TP+FN} \end{aligned}$$

(6)

$$\begin{aligned} F1Score = 2*\frac{Precision*Recall}{Precision+Recall} \end{aligned}$$

(7)