Opening the black box: interpretability of machine learning algorithms in electrocardiography

Developing and deploying deep learning models in brain magnetic resonance imaging: A review

Magnetic resonance imaging (MRI) of the brain has benefited from deep learning (DL) to alleviate the burden on radiologists and MR technologists, and improve throughput. The easy accessibility of DL tools has resulted in a rapid increase of DL models and subsequent peer-reviewed publications. However, the rate of deployment in clinical settings is low. Therefore, this review attempts to bring together the ideas from data collection to deployment in the clinic, building on the guidelines and principles that accreditation agencies have espoused. We introduce the need for and the role of DL to deliver accessible MRI. This is followed by a brief review of DL examples in the context of neuropathologies. Based on these studies and others, we collate the prerequisites to develop and deploy DL models for brain MRI. We then delve into the guiding principles to develop good machine learning practices in the context of neuroimaging, with a focus on explainability. A checklist based on the United States Food and Drug Administration's good machine learning practices is provided as a summary of these guidelines. Finally, we review the current challenges and future opportunities in DL for brain MRI.

Machine learning derived ECG risk score improves cardiovascular risk assessment in conjunction with coronary artery calcium scoring

Background

Precision estimation of cardiovascular risk remains the cornerstone of atherosclerotic cardiovascular disease (ASCVD) prevention. While coronary artery calcium (CAC) scoring is the best available non-invasive quantitative modality to evaluate risk of ASCVD, it excludes risk related to prior myocardial infarction, cardiomyopathy, and arrhythmia which are implicated in ASCVD. The high-dimensional and inter-correlated nature of ECG data makes it a good candidate for analysis using machine learning techniques and may provide additional prognostic information not captured by CAC. In this study, we aimed to develop a quantitative ECG risk score (eRiS) to predict major adverse cardiovascular events (MACE) alone, or when added to CAC. Further, we aimed to construct and validate a novel nomogram incorporating ECG, CAC and clinical factors for ASCVD.

Methods

We analyzed 5,864 patients with at least 1 cardiovascular risk factor who underwent CAC scoring and a standard ECG as part of the CLARIFY study (ClinicalTrials.gov Identifier: NCT04075162). Events were defined as myocardial infarction, coronary revascularization, stroke or death. A total of 649 ECG features, consisting of measurements such as amplitude and interval measurements from all deflections in the ECG waveform (53 per lead and 13 overall) were automatically extracted using a clinical software (GE Muse™ Cardiology Information System, GE Healthcare). The data was split into 4 training (S_tr) and internal validation (S_v) sets [S_tr (1): S_v (1): 50:50; S_tr (2): S_v (2): 60:40; S_tr (3): S_v (3): 70:30; S_tr (4): S_v (4): 80:20], and the results were compared across all the subsets. We used the ECG features derived from S_tr to develop eRiS. A least absolute shrinkage and selection operator-Cox (LASSO-Cox) regularization model was used for data dimension reduction, feature selection, and eRiS construction. A Cox-proportional hazards model was used to assess the benefit of using an eRiS alone (M_ecg), CAC alone (M_cac) and a combination of eRiS and CAC (M_ecg+cac) for MACE prediction. A nomogram (M_nom) was further constructed by integrating eRiS with CAC and demographics (age and sex). The primary endpoint of the study was the assessment of the performance of M_ecg, M_cac, M_ecg+cac and M_nom in predicting CV disease-free survival in ASCVD.

Findings

Over a median follow-up of 14 months, 494 patients had MACE. The feature selection strategy preserved only about 18% of the features that were consistent across the various strata (S_tr). The M_ecg model, comprising of eRiS alone was found to be significantly associated with MACE and had good discrimination of MACE (C-Index: 0.7, p = <2e-16). eRiS could predict time-to MACE (C-Index: 0.6, p = <2e-16 across all S_v). The M_ecg+cac model was associated with MACE (C-index: 0.71). Model comparison showed that M_ecg+cac was superior to M_ecg (p = 1.8e-10) or M_cac (p < 2.2e-16) alone. The M_nom, comprising of eRiS, CAC, age and sex was associated with MACE (C-index 0.71). eRiS had the most significant contribution, followed by CAC score and other clinical variables. Further, M_nom was able to identify unique patient risk-groups based on eRiS, CAC and clinical variables.

Conclusion

The use of ECG features in conjunction with CAC may allow for improved prognostication and identification of populations at risk. Future directions will involve prospective validation of the risk score and the nomogram across diverse populations with a heterogeneity of treatment effects.

Explainable AI Points to White Matter Hyperintensities for Alzheimer's Disease Identification: a Preliminary Study

Deep Learning approaches are powerful tools in a great variety of classification tasks. However, they are limitedly accepted or trusted in clinical frameworks due to their typical "black box" outline: their architecture is well-known, but processes employed in classification are often inaccessible to humans. With this work, we explored the problem of "Explainable AI" (XAI) in Alzheimer's disease (AD) classification tasks. Data from a neuroimaging cohort (n = 251 from OASIS-3) of early-stage AD dementia and healthy controls (HC) were analysed. The MR scans were initially fed to a pre-trained DL model, which achieved good performance on the test set (AUC: 0.82, TPR: 0.78, TNR: 0.81). Results were then investigated by means of an XAI approach (Occlusion Sensitivity method) that provided measures of relevance (RV) as outcome. We compared RV values obtained within healthy tissues with those underlying white matter hyperintensity (WMH) lesions. The analysis was conducted on 4 different groups of data, obtained by stratifying correct and misclassified images according to the health condition of participants (AD/HC). Results highlighted that the DL model found favourable leveraging lesioned brain areas for AD identification. A statistically significant difference ( ) between WMH and healthy tissue contributions was indeed observed for AD recognition, differently from the HC case ( p=0.27). Clinical Relevance - This study, though preliminary, suggested that DL models might be trained to use known clinical information and reinforced the role of WMHs as neuroimaging biomarker for AD dementia. The outlined findings have a significant clinical relevance as they prepare the ground for a progressive increase in the level of trust laid in DL approaches.

Clinical significance, challenges and limitations in using artificial intelligence for electrocardiography-based diagnosis

Cardiovascular diseases are one of the leading global causes of mortality. Currently, clinicians rely on their own analyses or automated analyses of the electrocardiogram (ECG) to obtain a diagnosis. However, both approaches can only include a finite number of predictors and are unable to execute complex analyses. Artificial intelligence (AI) has enabled the introduction of machine and deep learning algorithms to compensate for the existing limitations of current ECG analysis methods, with promising results. However, it should be prudent to recognize that these algorithms also associated with their own unique set of challenges and limitations, such as professional liability, systematic bias, surveillance, cybersecurity, as well as technical and logistical challenges. This review aims to increase familiarity with and awareness of AI algorithms used in ECG diagnosis, and to ultimately inform the interested stakeholders on their potential utility in addressing present clinical challenges.

Advanced computation in cardiovascular physiology: new challenges and opportunities

Recent developments in computational physiology have successfully exploited advanced signal processing and artificial intelligence tools for predicting or uncovering characteristic features of physiological and pathological states in humans. While these advanced tools have demonstrated excellent diagnostic capabilities, the high complexity of these computational 'black boxes' may severely limit scientific inference, especially in terms of biological insight about both physiology and pathological aberrations. This theme issue highlights current challenges and opportunities of advanced computational tools for processing dynamical data reflecting autonomic nervous system dynamics, with a specific focus on cardiovascular control physiology and pathology. This includes the development and adaptation of complex signal processing methods, multivariate cardiovascular models, multiscale and nonlinear models for central-peripheral dynamics, as well as deep and transfer learning algorithms applied to large datasets. The width of this perspective highlights the issues of specificity in heartbeat-related features and supports the need for an imminent transition from the black-box paradigm to explainable and personalized clinical models in cardiovascular research. This article is part of the theme issue 'Advanced computation in cardiovascular physiology: new challenges and opportunities'.