Demonstration of the potential of white-box machine learning approaches to gain insights from cardiovascular disease electrocardiograms

Abstract

We present the results from a white-box machine learning approach to detect cardiac arrhythmias using electrocardiographic data. A C5.0 is trained to recognize four classes using common features. The four classes are (i) atrial fibrillation and atrial flutter, (ii) tachycardias (iii), sinus bradycardia and (iv) sinus rhythm. Data from 10,646 subjects, 83% of whom have at least one arrhythmia and 17% of whom exhibit a normal sinus rhythm, are used. The C5.0 is trained using 10-fold cross-validation and is able to achieve a balanced accuracy of 95.35%. By using the white-box machine learning approach, a clear and comprehensible tree structure can be revealed, which has selected the 5 most important features from a total of 24 features. These 5 features are ventricular rate, RR-Interval variation, atrial rate, age and difference between longest and shortest RR-Interval. The combination of ventricular rate, RR-Interval variation and atrial rate is especially relevant to achieve classification accuracy, which can be disclosed through the tree. The tree assigns unique values to distinguish the classes. These findings could be applied in medicine in the future. It can be shown that a white-box machine learning approach can reveal granular structures, thus confirming known linear relationships and also revealing nonlinear relationships. To highlight the strength of the C5.0 with respect to this structural revelation, the results of further white-box machine learning and black-box machine learning algorithms are presented.

Addressing bias in big data and AI for health care: A call for open science

“…As an example, transparent techniques like decision trees, relying on interpretable features, can provide a “white-box” approach for diagnosis. 68 …”

“… 69 For example, the kernels or intermediate features of a trained neural network may shed light on the learned structure in different layers of the network, giving rise to methods like class activation mapping (CAM). 67 Other methods are gradient based, like saliency maps, 68 and calculate the contribution of each input pixel to the overall classification performance. The combination of these two approaches has given rise to Grad-CAM, 70 which allows the identification of regions of interest in the input data that mostly influenced the network’s decision.…”

Detecting atrial fibrillation with a wearable device

“…For the SAFE dataset the algorithm correctly predicts all of the signals to have SR. For the Chapman University dataset with more challenging rhythms, the performance metrics are predictably lower but still very high with an accuracy measure of 95.3%. For reference, one study working on the same database was able to achieve a very similar result with the accuracy of 95.35%, but with much more complicated machine learning system [8]. Figure 4 demonstrates the confusion matrix for our results of the Chapman University database and gives more insight into the accuracy of the algorithm.…”

Logic-based mechanistic machine learning on high-content images reveals how drugs differentially regulate cardiac fibroblasts

“…We determined drug concentrations via a literature search, prioritizing concentrations that yielded significant effects in vitro in fibroblasts or similar cell types. The drugs with their respective concentrations are as follows: [0.25,1,2] µg/ml of anakinra (Kineret, SOBI Inc.), [1,5,10] µM valsartan (Sigma-Aldrich, SML0142-10MG), [0.2,1,2] µM BNP (Sigma-Aldrich, B5900-.5MG), [1,5,10]µM valsartan combos respectively with [0.2,1,2] µM BNP, [10,30,60]mM glutathione (Sigma-Aldrich, G4251-1G), [1,3,5] µM CW-HM12 (Cayman Chemical Company, 19480), [10,20,50] µM salbutamol (Sigma-Aldrich, S8260-25MG), [5,10,25] µM marimistat (Sigma-Aldrich, M2699-5Mg), [1,5,10] µM galunisertib (Selleck Chemicals, S2230), [12.5,25,50] µM fasudil (Sigma-Aldrich, CDS021620-10MG), [10,25,50 ]µM SB203580 (Sigma-Aldrich, S8307-1MG), [1,5,10] mg/mL pirfenidone (Sigma-Aldrich, P2116-10MG), [5,10,20] µM defactinib (MedChem Express, HY-12289A), [5,10,20] µM WH-4-023 (Sigma-Aldrich, SML1334-5MG), and 20 µM LY294002 (Selleck Chemicals, S1105). Cells were grown in these conditions for 72 hours.…”