Machine learning promises to assist physicians with predictions of mortality and of other future clinical events by learning complex patterns from historical data, such as longitudinal electronic ...health records. Here we show that a convolutional neural network trained on raw pixel data in 812,278 echocardiographic videos from 34,362 individuals provides superior predictions of one-year all-cause mortality. The model's predictions outperformed the widely used pooled cohort equations, the Seattle Heart Failure score (measured in an independent dataset of 2,404 patients with heart failure who underwent 3,384 echocardiograms), and a machine learning model involving 58 human-derived variables from echocardiograms and 100 clinical variables derived from electronic health records. We also show that cardiologists assisted by the model substantially improved the sensitivity of their predictions of one-year all-cause mortality by 13% while maintaining prediction specificity. Large unstructured datasets may enable deep learning to improve a wide range of clinical prediction models.
Cardiovascular magnetic resonance using displacement encoding with stimulated echoes (DENSE) is capable of assessing advanced measures of cardiac mechanics such as strain and torsion. A potential ...hurdle to widespread clinical adoption of DENSE is the time required to manually segment the myocardium during post-processing of the images. To overcome this hurdle, we proposed a radical approach in which only three contours per image slice are required for post-processing (instead of the typical 30-40 contours per image slice). We hypothesized that peak left ventricular circumferential, longitudinal and radial strains and torsion could be accurately quantified using this simplified analysis.
We tested our hypothesis on a large multi-institutional dataset consisting of 541 DENSE image slices from 135 mice and 234 DENSE image slices from 62 humans. We compared measures of cardiac mechanics derived from the simplified post-processing to those derived from original post-processing utilizing the full set of 30-40 manually-defined contours per image slice. Accuracy was assessed with Bland-Altman limits of agreement and summarized with a modified coefficient of variation. The simplified technique showed high accuracy with all coefficients of variation less than 10% in humans and 6% in mice. The accuracy of the simplified technique was also superior to two previously published semi-automated analysis techniques for DENSE post-processing.
Accurate measures of cardiac mechanics can be derived from DENSE cardiac magnetic resonance in both humans and mice using a simplified technique to reduce post-processing time by approximately 94%. These findings demonstrate that quantifying cardiac mechanics from DENSE data is simple enough to be integrated into the clinical workflow.
Displacement Encoding with Stimulated Echoes (DENSE) encodes displacement into the phase of the magnetic resonance signal. Due to the stimulated echo, the signal is inherently low and fades through ...the cardiac cycle. To compensate, a spiral acquisition has been used at 1.5T. This spiral sequence has not been validated at 3T, where the increased signal would be valuable, but field inhomogeneities may result in measurement errors. We hypothesized that spiral cine DENSE is valid at 3T and tested this hypothesis by measuring displacement errors at both 1.5T and 3T in vivo.
Two-dimensional spiral cine DENSE and tagged imaging of the left ventricle were performed on ten healthy subjects at 3T and six healthy subjects at 1.5T. Intersection points were identified on tagged images near end-systole. Displacements from the DENSE images were used to project those points back to their origins. The deviation from a perfect grid was used as a measure of accuracy and quantified as root-mean-squared error. This measure was compared between 3T and 1.5T with the Wilcoxon rank sum test. Inter-observer variability of strains and torsion quantified by DENSE and agreement between DENSE and harmonic phase (HARP) were assessed by Bland-Altman analyses. The signal to noise ratio (SNR) at each cardiac phase was compared between 3T and 1.5T with the Wilcoxon rank sum test.
The displacement accuracy of spiral cine DENSE was not different between 3T and 1.5T (1.2 ± 0.3 mm and 1.2 ± 0.4 mm, respectively). Both values were lower than the DENSE pixel spacing of 2.8 mm. There were no substantial differences in inter-observer variability of DENSE or agreement of DENSE and HARP between 3T and 1.5T. Relative to 1.5T, the SNR at 3T was greater by a factor of 1.4 ± 0.3.
The spiral cine DENSE acquisition that has been used at 1.5T to measure cardiac displacements can be applied at 3T with equivalent accuracy. The inter-observer variability and agreement of DENSE-derived peak strains and torsion with HARP is also comparable at both field strengths. Future studies with spiral cine DENSE may take advantage of the additional SNR at 3T.
Displacement encoding with stimulated echoes (DENSE) is a phase contrast technique that encodes tissue displacement into phase images, which are typically processed into measures of cardiac function ...such as strains. For improved signal to noise ratio and spatiotemporal resolution, DENSE is often acquired with a spiral readout using an 11.1 ms readout duration. However, long spiral readout durations are prone to blurring due to common phenomena such as off-resonance and T2* decay, which may alter the resulting quantifications of strain. We hypothesized that longer readout durations would reduce image quality and underestimate cardiac strains at both 3.0 T and 1.5 T and that using short readout durations could overcome these limitations.
Computational simulations were performed to investigate the relationship between off-resonance and T2* decay, the spiral cine DENSE readout duration, and measured radial and circumferential strain. Five healthy participants subsequently underwent 2D spiral cine DENSE at both 3.0 T and 1.5 T with several different readout durations 11.1 ms and shorter. Pearson correlations were used to assess the relationship between cardiac strains and the spiral readout duration.
Simulations demonstrated that long readout durations combined with off-resonance and T2* decay yield blurred images and underestimate strains. With the typical 11.1 ms DENSE readout, blurring was present in the anterior and lateral left ventricular segments of participants and was markedly improved with shorter readout durations. Radial and circumferential strains from those segments were significantly correlated with the readout duration. Compared to the 1.9 ms readout, the 11.1 ms readout underestimated radial and circumferential strains in those segments at both field strengths by up to 19.6% and 1.5% (absolute), or 42% and 7% (relative), respectively.
Blurring is present in spiral cine DENSE images acquired at both 3.0 T and 1.5 T using the typical 11.1 ms readout duration, which yielded substantially reduced radial strains and mildly reduced circumferential strains. Clinical studies using spiral cine DENSE should consider these limitations, while future technical advances may need to leverage accelerated techniques to improve the robustness and accuracy of the DENSE acquisition rather than focusing solely on reduced acquisition time.
Rat models have assumed an increasingly important role in cardiac research. However, a detailed profile of regional cardiac mechanics, such as strains and torsion, is lacking for rats. We ...hypothesized that healthy rat left ventricles (LVs) exhibit regional differences in cardiac mechanics, which are part of normal function. In this study, images of the LV were obtained with 3D cine displacement encoding with stimulated echoes (DENSE) cardiovascular magnetic resonance in 10 healthy rats. To evaluate regional cardiac mechanics, the LV was divided into basal, mid‐ventricular, and apical regions. The myocardium at the mid‐LV was further partitioned into four wall segments (i.e. septal, inferior, lateral, and anterior) and three transmural layers (i.e. sub‐endocardium, mid‐myocardium, and sub‐epicardium). The six Lagrangian strain components (i.e. Err, Ecc, Ell, Ecl, Erl, and Ecr) were computed from the 3D displacement field and averaged within each region of interest. Torsion was quantified using the circumferential‐longitudinal shear angle. While peak systolic Ecl differed between the mid‐ventricle and apex, the other five components of peak systolic strain were similar across the base, mid‐ventricle, and apex. In the mid‐LV myocardium, Ecc decreased gradually from the sub‐endocardial to the sub‐epicardial layer. Ell demonstrated significant differences between the four wall segments, with the largest magnitude in the inferior segment. Err was uniform among the four wall segments. Ecl varied along the transmural direction and among wall segments, whereas Erl differed only among the wall segments. Erc was not associated with significant variations. Torsion also varied along the transmural direction and among wall segments. These results provide fundamental insights into the regional contractile function of healthy rat hearts, and form the foundation for future studies on regional changes induced by disease or treatments.
Based on 3D cine DENSE measurements, peak systolic Ecc and Ell varied among the three transmural layers (i.e. sub‐endocardium, mid‐myocardium, and sub‐epicardium) and among the four wall segments (i.e. septal, inferior, lateral, and anterior), respectively, at the mid‐ventricle of healthy rats. Peak systolic Ecl and related ventricular torsion exhibited notable differences along the transmural direction and among wall segments.
Left ventricular (LV) torsion is an important indicator of cardiac function that is limited by high inter-test variability (50% of the mean value). We hypothesized that this high inter-test ...variability is partly due to inconsistent breath-hold positions during serial image acquisitions, which could be significantly improved by using a respiratory navigator for cardiovascular magnetic resonance (CMR) based quantification of LV torsion.
We assessed respiratory-related variability in measured LV torsion with two distinct experimental protocols. First, 17 volunteers were recruited for CMR with cine displacement encoding with stimulated echoes (DENSE) in which a respiratory navigator was used to measure and then enforce variability in end-expiratory position between all LV basal and apical acquisitions. From these data, we quantified the inter-test variability of torsion in the absence and presence of enforced end-expiratory position variability, which established an upper bound for the expected torsion variability. For the second experiment (in 20 new, healthy volunteers), 10 pairs of cine DENSE basal and apical images were each acquired from consecutive breath-holds and consecutive navigator-gated scans (with a single acceptance position). Inter-test variability of torsion was compared between the breath-hold and navigator-gated scans to quantify the variability due to natural breath-hold variation. To demonstrate the importance of these variability reductions, we quantified the reduction in sample size required to detect a clinically meaningful change in LV torsion with the use of a respiratory navigator.
The mean torsion was 3.4 ± 0.2°/cm. From the first experiment, enforced variability in end-expiratory position translated to considerable variability in measured torsion (0.56 ± 0.34°/cm), whereas inter-test variability with consistent end-expiratory position was 57% lower (0.24 ± 0.16°/cm, p < 0.001). From the second experiment, natural respiratory variability from consecutive breath-holds translated to a variability in torsion of 0.24 ± 0.10°/cm, which was significantly higher than the variability from navigator-gated scans (0.18 ± 0.06°/cm, p = 0.02). By using a respiratory navigator with DENSE, theoretical sample sizes were reduced from 66 to 16 and 26 to 15 as calculated from the two experiments.
A substantial portion (22-57%) of the inter-test variability of LV torsion can be reduced by using a respiratory navigator to ensure a consistent breath-hold position between image acquisitions.
The goal of this study was to use machine learning to more accurately predict survival after echocardiography.
Predicting patient outcomes (e.g., survival) following echocardiography is primarily ...based on ejection fraction (EF) and comorbidities. However, there may be significant predictive information within additional echocardiography-derived measurements combined with clinical electronic health record data.
Mortality was studied in 171,510 unselected patients who underwent 331,317 echocardiograms in a large regional health system. The authors investigated the predictive performance of nonlinear machine learning models compared with that of linear logistic regression models using 3 different inputs: 1) clinical variables, including 90 cardiovascular-relevant International Classification of Diseases, Tenth Revision, codes, and age, sex, height, weight, heart rate, blood pressures, low-density lipoprotein, high-density lipoprotein, and smoking; 2) clinical variables plus physician-reported EF; and 3) clinical variables and EF, plus 57 additional echocardiographic measurements. Missing data were imputed with a multivariate imputation by using a chained equations algorithm (MICE). The authors compared models versus each other and baseline clinical scoring systems by using a mean area under the curve (AUC) over 10 cross-validation folds and across 10 survival durations (6 to 60 months).
Machine learning models achieved significantly higher prediction accuracy (all AUC >0.82) over common clinical risk scores (AUC = 0.61 to 0.79), with the nonlinear random forest models outperforming logistic regression (p < 0.01). The random forest model including all echocardiographic measurements yielded the highest prediction accuracy (p < 0.01 across all models and survival durations). Only 10 variables were needed to achieve 96% of the maximum prediction accuracy, with 6 of these variables being derived from echocardiography. Tricuspid regurgitation velocity was more predictive of survival than LVEF. In a subset of studies with complete data for the top 10 variables, multivariate imputation by chained equations yielded slightly reduced predictive accuracies (difference in AUC of 0.003) compared with the original data.
Machine learning can fully utilize large combinations of disparate input variables to predict survival after echocardiography with superior accuracy.
Display omitted
Purpose
To determine the optimal respiratory navigator gating configuration for the quantification of left ventricular strain using spiral cine displacement encoding with stimulated echoes (DENSE) ...MRI.
Materials and Methods
Two‐dimensional spiral cine DENSE was performed on a 3 Tesla MRI using two single‐navigator configurations (retrospective, prospective) and a combined “dual‐navigator” configuration in 10 healthy adults and 20 healthy children. The adults also underwent breathhold DENSE as a reference standard for comparisons. Peak left ventricular strains, signal‐to‐noise ratio (SNR), and navigator efficiency were compared. Subjects also underwent dual‐navigator gating with and without visual feedback to determine the effect on navigator efficiency.
Results
There were no differences in circumferential, radial, and longitudinal strains between navigator‐gated and breathhold DENSE (P = 0.09–0.95) (as confidence intervals, retrospective: ‐1.0%–1.1%, ‐7.4%–2.0%, ‐1.0%–1.2%; prospective: ‐0.6%–2.7%, ‐2.8%–8.3%, ‐0.3%–2.9%; dual: ‐1.6%–0.5%, ‐8.3%–3.2%, ‐0.8%–1.9%, respectively). The dual configuration maintained SNR compared with breathhold acquisitions (16 versus 18, P = 0.06). SNR for the prospective configuration was lower than for the dual navigator in adults (P = 0.004) and children (P < 0.001). Navigator efficiency was higher (P < 0.001) for both retrospective (54%) and prospective (56%) configurations compared with the dual configuration (35%). Visual feedback improved the dual configuration navigator efficiency to 55% (P < 0.001).
Conclusion
When quantifying left ventricular strains using spiral cine DENSE MRI, a dual navigator configuration results in the highest SNR in adults and children. In adults, a retrospective configuration has good navigator efficiency without a substantial drop in SNR. Prospective gating should be avoided because it has the lowest SNR. Visual feedback represents an effective option to maintain navigator efficiency while using a dual navigator configuration.
Level of Evidence: 2
J. Magn. Reson. Imaging 2017;45:786–794.
Advanced cardiovascular magnetic resonance (CMR) acquisitions often require long scan durations that necessitate respiratory navigator gating. The tradeoff of navigator gating is reduced scan ...efficiency, particularly when the patient's breathing patterns are inconsistent, as is commonly seen in children. We hypothesized that engaging pediatric participants with a navigator-controlled videogame to help control breathing patterns would improve navigator efficiency and maintain image quality.
We developed custom software that processed the Siemens respiratory navigator image in real-time during CMR and represented diaphragm position using a cartoon avatar, which was projected to the participant in the scanner as visual feedback. The game incentivized children to breathe such that the avatar was positioned within the navigator acceptance window (±3 mm) throughout image acquisition. Using a 3T Siemens Tim Trio, 50 children (Age: 14 ± 3 years, 48 % female) with no significant past medical history underwent a respiratory navigator-gated 2D spiral cine displacement encoding with stimulated echoes (DENSE) CMR acquisition first with no feedback (NF) and then with the feedback game (FG). Thirty of the 50 children were randomized to undergo extensive off-scanner training with the FG using a MRI simulator, or no off-scanner training. Navigator efficiency, signal-to-noise ratio (SNR), and global left-ventricular strains were determined for each participant and compared.
Using the FG improved average navigator efficiency from 33 ± 15 to 58 ± 13 % (p < 0.001) and improved SNR by 5 % (p = 0.01) compared to acquisitions with NF. There was no difference in navigator efficiency (p = 0.90) or SNR (p = 0.77) between untrained and trained participants for FG acquisitions. Circumferential and radial strains derived from FG acquisitions were slightly reduced compared to NF acquisitions (-16 ± 2 % vs -17 ± 2 %, p < 0.001; 40 ± 10 % vs 44 ± 11 %, p = 0.005, respectively). There were no differences in longitudinal strain (p = 0.38).
Use of a respiratory navigator feedback game during navigator-gated CMR improved navigator efficiency in children from 33 to 58 %. This improved efficiency was associated with a 5 % increase in SNR for spiral cine DENSE. Extensive off-scanner training was not required to achieve the improvement in navigator efficiency.
Weather and climate extremes have been varying and changing on many different time scales. In recent decades, heat waves have generally become more frequent across the United States, while cold waves ...have been decreasing. While this is in keeping with expectations in a warming climate, it turns out that decadal variations in the number of U.S. heat and cold waves do not correlate well with the observed U.S. warming during the last century. Annual peak flow data reveal that river flooding trends on the century scale do not show uniform changes across the country. While flood magnitudes in the Southwest have been decreasing, flood magnitudes in the Northeast and north-central United States have been increasing. Confounding the analysis of trends in river flooding is multiyear and even multidecadal variability likely caused by both large-scale atmospheric circulation changes and basin-scale “memory” in the form of soil moisture. Droughts also have long-term trends as well as multiyear and decadal variability. Instrumental data indicate that the Dust Bowl of the 1930s and the drought in the 1950s were the most significant twentieth-century droughts in the United States, while tree ring data indicate that the megadroughts over the twelfth century exceeded anything in the twentieth century in both spatial extent and duration. The state of knowledge of the factors that cause heat waves, cold waves, floods, and drought to change is fairly good with heat waves being the best understood.