Meeting the unmet needs of clinicians from AI systems showcased for cardiology with deep-learning-based ECG analysis

Abstract

Despite their great promise, artificial intelligence (AI) systems have yet to become ubiquitous in the daily practice of medicine largely due to several crucial unmet needs of healthcare practitioners. These include lack of explanations in clinically meaningful terms, handling the presence of unknown medical conditions, and transparency regarding the system's limitations, both in terms of statistical performance as well as recognizing situations for which the system's predictions are irrelevant. We articulate these unmet clinical needs as machine-learning (ML) problems and systematically address them with cutting-edge ML techniques. We focus on electrocardiogram (ECG) analysis as an example domain in which AI has great potential and tackle two challenging tasks: the detection of a heterogeneous mix of known and unknown arrhythmias from ECG and the identification of underlying cardio-pathology from segments annotated as normal sinus rhythm recorded in patients with an intermittent arrhythmia. We validate our methods by simulating a screening for arrhythmias in a large-scale population while adhering to statistical significance requirements. Specifically, our system 1) visualizes the relative importance of each part of an ECG segment for the final model decision; 2) upholds specified statistical constraints on its out-of-sample performance and provides uncertainty estimation for its predictions; 3) handles inputs containing unknown rhythm types; and 4) handles data from unseen patients while also flagging cases in which the model's outputs are not usable for a specific patient. This work represents a significant step toward overcoming the limitations currently impeding the integration of AI into clinical practice in cardiology and medicine in general.

Keywords: artificial intelligence; cardiology; deep learning; medical.

Fig. 1.

Schematic of our framework. (A) Multiple ECG leads, possibly containing a mix of known and unknown rhythm types, are provided as inputs for the model. (B) Our custom STA layer is applied, producing input importance maps highlighting the regions of input contributing the most to the prediction. (C) A deep neural network analyzes the attended inputs. (D) Separate binary classifiers provide the probability of each rhythm type; different classification thresholds are used for different rhythms based on statistical significance requirements together with a distribution indicator function that expresses whether the model’s predictions are relevant for the given input. (E) Finally, a clinician chooses whether to trust the model based on the indicator value and uses the rhythm predictions combined with the STA-highlighted regions to make an AI-supported clinical decision. Note that the signals portrayed are only schematic and are not physiological recordings.

Fig. 2.

Temporal attention maps generated by the model for four test set samples, each containing two ECG leads and belonging to a different patient. For clarity, only 9 s are displayed from each sample. (A) NSR from a healthy subject with no underlying cardiac pathology. (B) NSR from a patient with an existing underlying cardiac pathology (LP-NSR). (C) AF rhythm. (D) VT rhythm. Blue arrows denote morphologically normal features in normal sinus segments, red arrows denote noise, and black arrows denote abnormal features in ECG segments related to an arrhythmia. These examples showcase the ability of the STA mechanism to detect and highlight periodic components in the input due to the way it is calculated from its spectral representation. Therefore, note that the arrows shown were chosen to present a few exemplary features of interest; they do not represent all relevant high-attention morphological features.

Fig. 3.

Multiclass prediction. An input sample to the model is a 30- or 60-s ECG segment, which may simultaneously contain zero or more known types of ECG rhythms as well as unknown types. All the known rhythm types present in the same sample comprise the set of its ground-truth labels. The model outputs a likelihood estimation for each known rhythm type (blue or red bars) along with an uncertainty estimation, which is equivalent to 1 SD in the prediction probability (vertical black lines). A positive prediction for a specific rhythm (filled blue bar) is produced if its likelihood value crosses a predetermined threshold, and a negative prediction is produced otherwise (empty red bar). For visualization only, the threshold was set to the same value of 0.9 for all rhythm types (horizontal black line), though it should generally be chosen per rhythm type according to statistical requirements. Four samples are shown containing one or more known or unknown rhythm types: (A) sample containing only one known rhythm (AF); (B) sample containing two known rhythms (LP-NSR and ventricular bigeminy); (C) sample containing one known (AF) and one unknown (Paced) rhythm, hence there is no output for the unknown type; and (D) sample with a single unknown rhythm type (Paced). supraventricular tachycardia, SVT; ventricular bigeminy, Vent. Big.; ventricular trigeminy, Vent. Trig.; idioventricular rhythm, IR; atrial bigeminy, At. Big.; sinus bradycardia, Brady.