At a Young Investigator Award Session today, Mr. Dimitrios Doudesis (University of Edinburgh, UK) presented research on the validation of the myocardial-ischemic-injury-index (MI³) machine-learning model to predict myocardial infarction (MI) diagnosis. This promising machine-learning algorithm has been shown to predict the likelihood of MI in patients with suspected acute coronary syndrome (ACS);1 however, whether this algorithm performs well in unselected patients or predicts recurrent events was unknown.
Mr. Doudesis et al performed an observational analysis from a randomised trial funded by the British Heart Foundation, which included patients with suspected ACS and serial high-sensitivity cardiac troponin I measurements without ST-segment elevation MI. Using gradient boosting, MI³ incorporates age, sex and two troponin measurements to compute a value (0–100) reflecting an individual’s likelihood of MI, and estimates the negative predictive value (NPV) and positive predictive value (PPV). Model performance for an index diagnosis of MI and for subsequent MI or cardiovascular death at 1 year was determined using previously defined low- and high-probability thresholds (1.6 and 49.7, respectively)
Of 20,761 patients in the validation cohort, 3,278 (15.8%) had MI. MI³ accurately estimated the likelihood of MI with an area under the receiver-operating-characteristic curve (AUC) of 0.949 (95% confidence interval 0.946–0.952). In total, 62.5% of patients were identified as low-probability (sensitivity 99.3% [99.0–99.6%], NPV 99.8% [99.8–99.9%]) and 14.3% as high-probability (specificity 95.0% [94.7–95.3%], PPV 70.4% [69.0–71.9%]). At 1 year, MI³ accurately predicted the probability of subsequent adverse cardiovascular events, which were found to occur more often in high vs. low-probability patients (17.6% vs. 1.5%; p<0.001).
Accurate outcomes prediction with machine learning was also demonstrated in an ePoster by Doctor Lisa Tang and colleagues from the University of British Columbia (Vancouver, Canada). In this study, daily heart rhythm data were used to predict outcomes after ablation for atrial fibrillation (AF). Data were analysed from the CIRCA-DOSE randomised clinical trial that recorded daily measures, including heart rate variability and AF burden (total minutes in AF/day) from 343 patients referred for first catheter ablation due to AF refractory to at least one anti-arrhythmic drug. Comparative analyses were performed on 19 candidate model variants (e.g. stepwise regression, random forest, linear discriminant analysis etc.) to evaluate each model’s ability to identify patients predicted to experience at least one episode of AF recurrence post-ablation. They also examined the use of baseline comorbidities, echocardiographic parameters (e.g. left atrial volume) and CHA2DS2VASc scores. Multiple sets of threefold cross-validation experiments were conducted, each fold independently trained a candidate model with two thirds of the dataset and performed evaluation on the left-out samples.
Cross-validation results demonstrated that linear discriminant analysis and random forest models gave comparable performances, with a model based on random forest achieving the highest AUC of 0.68 ± 0.06 when 30 days of pre-ablation rhythm data were used (sensitivity of 65.9 ± 7.82; specificity of 66.3 ± 0.57). In contrast, the use of left atrial volume alone was not adequate to predict outcome (AUC ~0.5), and when combined with other baseline variables, the best model achieved an AUC of 0.58 ± 0.05. Feature analyses from the trained models suggest that pre-ablation AF burden had the highest relevance in predicting outcome. Using only daily AF burden, random forest and linear discriminant analysis models achieved AUCs of 0.61 ± 0.04 and 0.65 ± 0.04, respectively. These results indicate the value of pre-ablation rhythm data and machine learning for improving outcome prediction. Future analyses are planned to validate these findings using independent randomised trial data.
Image acquisition is another area where machine learning shows great promise. Recently, a deep-learning algorithm (Caption Guidance, Caption Health, Brisbane, CA, USA) was developed to guide the acquisition of standard transthoracic echocardiographic (TTE) views, including by novice users, which may be useful in clinical settings where immediate interrogation of anatomy is needed and in settings with limited resources. A ‘Best ePoster’ first-authored by Mr. Sam Surette (Caption Health, Brisbane, CA, USA) further validated this software through a descriptive subgroup analysis of the pivotal study used to support US FDA authorisation. The researchers evaluated whether novice-acquired TTE images guided by the software were of diagnostic quality in patients with and without implanted electrophysiological (EP) devices, focusing on right ventricular (RV) size and function, which were thought to be sensitive to EP devices.
Using the deep-learning algorithm, 8 nurses without prior experience in echocardiography each acquired limited TTEs (10 views) from 30 patients. In addition, 235 patients were also scanned by a trained sonographer without software assistance. Five Level 3 echocardiographers independently assessed the diagnostic quality of the TTEs acquired by the nurses and sonographers to assess the effect of EP devices on software performance.
The key finding was that, using the deep-learning software, nurses were able to acquire TTEs of sufficient quality to make qualitative assessments of RV size and function in greater than 80% of cases for patients with and without implanted EP devices. There was no significant difference observed between nurse and sonographer-acquired scans, indicating that the software performance is robust to the presence of EP device leads visible in the images.