A deep neural network learning algorithm outperforms a conventional algorithm for emergency department electrocardiogram interpretation.

BACKGROUND: Cardiologs® has developed the first electrocardiogram (ECG) algorithm that uses a deep neural network (DNN) for full 12­lead ECG analysis, including rhythm, QRS and ST-T-U waves. We compared the accuracy of the first version of Cardiologs® DNN algorithm to the Mortara/Veritas® conventional algorithm in emergency department (ED) ECGs. METHODS: Individual ECG diagnoses were prospectively mapped to one of 16 pre-specified groups of ECG diagnoses, which were further classified as "major" ECG abnormality or not. Automated interpretations were compared to blinded experts'. The primary outcome was the performance of the algorithms in finding at least one "major" abnormality. The secondary outcome was the proportion of all ECGs for which all groups were identified, with no false negative or false positive groups ("accurate ECG interpretation"). Additionally, we measured sensitivity and positive predictive value (PPV) for any abnormal group. RESULTS: Cardiologs® vs. Veritas® accuracy for finding a major abnormality was 92.2% vs. 87.2% (p < 0.0001), with comparable sensitivity (88.7% vs. 92.0%, p = 0.086), improved specificity (94.0% vs. 84.7%, p < 0.0001) and improved positive predictive value (PPV 88.2% vs. 75.4%, p < 0.0001). Cardiologs® had accurate ECG interpretation for 72.0% (95% CI: 69.6-74.2) of ECGs vs. 59.8% (57.3-62.3) for Veritas® (P < 0.0001). Sensitivity for any abnormal group for Cardiologs® and Veritas®, respectively, was 69.6% (95CI 66.7-72.3) vs. 68.3% (95CI 65.3-71.1) (NS). Positive Predictive Value was 74.0% (71.1-76.7) for Cardiologs® vs. 56.5% (53.7-59.3) for Veritas® (P < 0.0001). CONCLUSION: Cardiologs' DNN was more accurate and specific in identifying ECGs with at least one major abnormal group. It had a significantly higher rate of accurate ECG interpretation, with similar sensitivity and higher PPV.

Inspirational Adversarial Image Generation.

The task of image generation started receiving some attention from artists and designers, providing inspiration for new creations. However, exploiting the results of deep generative models such as Generative Adversarial Networks can be long and tedious given the lack of existing tools. In this work, we propose a simple strategy to inspire creators with new generations learned from a dataset of their choice, while providing some control over the output. We design a simple optimization method to find the optimal latent parameters corresponding to the closest generation to any input inspirational image. Specifically, we allow the generation given an inspirational image of the user's choosing by performing several optimization steps to recover optimal parameters from the model's latent space. We tested several exploration methods from classical gradient descents to gradient-free optimizers. Many gradient-free optimizers just need comparisons (better/worse than another image), so they can even be used without numerical criterion nor inspirational image, only with human preferences. Thus, by iterating on one's preferences we can make robust facial composite or fashion generation algorithms. Our results on four datasets of faces, fashion images, and textures show that satisfactory images are effectively retrieved in most cases.

A deep neural network for 12-lead electrocardiogram interpretation outperforms a conventional algorithm, and its physician overread, in the diagnosis of atrial fibrillation.

BACKGROUND: Automated electrocardiogram (ECG) interpretations may be erroneous, and lead to erroneous overreads, including for atrial fibrillation (AF). We compared the accuracy of the first version of a new deep neural network 12-Lead ECG algorithm (Cardiologs®) to the conventional Veritas algorithm in interpretation of AF. METHODS: 24,123 consecutive 12-lead ECGs recorded over 6 months were interpreted by 1) the Veritas® algorithm, 2) physicians who overread Veritas® (Veritas®â€¯+ physician), and 3) Cardiologs® algorithm. We randomly selected 500 out of 858 ECGs with a diagnosis of AF according to either algorithm, then compared the algorithms' interpretations, and Veritas®â€¯+ physician, with expert interpretation. To assess sensitivity for AF, we analyzed a separate database of 1473 randomly selected ECGs interpreted by both algorithms and by blinded experts. RESULTS: Among the 500 ECGs selected, 399 had a final classification of AF; 101 (20.2%) had ≥1 false positive automated interpretation. Accuracy of Cardiologs® (91.2%; CI: 82.4-94.4) was higher than Veritas® (80.2%; CI: 76.5-83.5) (p < 0.0001), and equal to Veritas®â€¯+ physician (90.0%, CI:87.1-92.3) (p = 0.12). When Veritas® was incorrect, accuracy of Veritas®â€¯+ physician was only 62% (CI 52-71); among those ECGs, Cardiologs® accuracy was 90% (CI: 82-94; p < 0.0001). The second database had 39 AF cases; sensitivity was 92% vs. 87% (p = 0.46) and specificity was 99.5% vs. 98.7% (p = 0.03) for Cardiologs® and Veritas® respectively. CONCLUSION: Cardiologs® 12-lead ECG algorithm improves the interpretation of atrial fibrillation.

Network, anatomical, and non-imaging measures for the prediction of ADHD diagnosis in individual subjects.

Brain imaging methods have long held promise as diagnostic aids for neuropsychiatric conditions with complex behavioral phenotypes such as Attention-Deficit/Hyperactivity Disorder. This promise has largely been unrealized, at least partly due to the heterogeneity of clinical populations and the small sample size of many studies. A large, multi-center dataset provided by the ADHD-200 Consortium affords new opportunities to test methods for individual diagnosis based on MRI-observable structural brain attributes and functional interactions observable from resting-state fMRI. In this study, we systematically calculated a large set of standard and new quantitative markers from individual subject datasets. These features (>12,000 per subject) consisted of local anatomical attributes such as cortical thickness and structure volumes, and both local and global resting-state network measures. Three methods were used to compute graphs representing interdependencies between activations in different brain areas, and a full set of network features was derived from each. Of these, features derived from the inverse of the time series covariance matrix, under an L1-norm regularization penalty, proved most powerful. Anatomical and network feature sets were used individually, and combined with non-imaging phenotypic features from each subject. Machine learning algorithms were used to rank attributes, and performance was assessed under cross-validation and on a separate test set of 168 subjects for a variety of feature set combinations. While non-imaging features gave highest performance in cross-validation, the addition of imaging features in sufficient numbers led to improved generalization to new data. Stratification by gender also proved to be a fruitful strategy to improve classifier performance. We describe the overall approach used, compare the predictive power of different classes of features, and describe the most impactful features in relation to the current literature.