Development of a machine learning model using electrocardiogram signals to improve acute pulmonary embolism screening

Abstract

Aims Clinical scoring systems for pulmonary embolism (PE) screening have low specificity and contribute to CT pulmonary angiogram (CTPA) overuse. We assessed whether deep learning models using an existing and routinely collected data modality, electrocardiogram (ECG) waveforms, can increase specificity for PE detection. Methods and Results We create a retrospective cohort of 21,183 patients at moderate- to high-suspicion of PE and associate 23,793 CTPAs (10.0% PE-positive) with 320,746 ECGs and encounter-level clinical data (demographics, comorbidities, vital signs, and labs). We develop three machine learning models to predict PE likelihood: an ECG model using only ECG waveform data, an EHR model using tabular clinical data, and a Fusion model integrating clinical data and an embedded representation of the ECG waveform. We find that a Fusion model (area under receiver-operating characteristic [AUROC] 0.81 ± 0.01) outperforms both the ECG model (AUROC 0.59 ± 0.01) and EHR model (AUROC 0.65 ± 0.01). On a sample of 100 patients from the test set, the Fusion model also achieves greater specificity (0.18) and performance (AUROC 0.84 ± 0.01) than four commonly evaluated clinical scores: Wells' Criteria, Revised Geneva Score, Pulmonary Embolism Rule-Out Criteria, and 4-Level Pulmonary Embolism Clinical Probability Score (AUROC 0.50-0.58, specificity 0.00-0.05). The model is superior to these scores on feature sensitivity analyses (AUROC 0.66 to 0.84) and achieves comparable performance across sex (AUROC 0.81) and racial/ethnic (AUROC 0.77 to 0.84) subgroups. Conclusion Synergistic deep learning of electrocardiogram waveforms with traditional clinical variables can increase the specificity of PE detection in patients at least at moderate suspicion for PE.

Machine Learning-Based Predictive Models for Patients with Venous Thromboembolism: A Systematic Review

“…Overall, 15 studies explored ML-based models to assist VTE diagnosis, either in the form of a pretest probability or to assist diagnosis after clinical presentation (Supplementary Table 4) [79][80][81][82][83][84][85][86][87][88][89][90][91][92][93] . Most of them do not describe clearly preprocessing steps, splitting/cross-validation, hyperparameters, and other performance metrics.…”

“…Overall, ML based prediction of VTE is limited so far, and current studies are sparse and problematic so further work is needed in that direction exploiting the advantage of big data. Somani et al 89 used an extensive source of data from the EHR, besides vital signs, lab and clinical data, deep learning-based embedding representation of ECG waveforms, as well as unstructured data from imaging reports. They also followed STROBE 94 and TRIPOD 67 guidelines for model development.…”

“…Nevertheless, if it is not performed properly data leakage can happen. In every case an unseen part of the dataset must be kept for internal validation purposes and this was not clear in many [28][29][30]34,[38][39][40][41]51,53,54,[56][57][58]61,62,64,77,84,89,126,127 of the reviewed studies. This review highlighted a significant problem in AI-related research, namely the lack of standardized reporting methods.…”

Multimodal Fusion Models for Pulmonary Embolism Mortality Prediction

“…Other studies have reported similar advantages for a wide range of medical applications from radiology 17 to oncology 18 . For the task of PE detection, both Huang et al 19 and Somani et al 20 proposed a multimodal fusion model combining EHR data with CTPA and ECG respectively. Most previous work has focused on approaches using only one of several possible fusion strategies.…”

Multimodal fusion models for pulmonary embolism mortality prediction

“…Promising results for PE detection have been shown based on imaging data alone, and are further improved when combined with additional modalities. Both Huang et al 19 and Somani et al 20 proposed a multimodal fusion model combining EHR data with CTPA and ECG respectively for the detection task. When designing a multimodal fusion solution, several models can be considered: Multimodal vision-language models trained with a contrastive objective 21 , 22 have enabled zero-shot adaptation to novel tasks, without the need for fine-tuning.…”