A machine learning framework for the evaluation of myocardial rotation in patients with noncompaction cardiomyopathy.

AIMS: Noncompaction cardiomyopathy (NCC) is considered a genetic cardiomyopathy with unknown pathophysiological mechanisms. We propose to evaluate echocardiographic predictors for rigid body rotation (RBR) in NCC using a machine learning (ML) based model. METHODS AND
RESULTS: Forty-nine outpatients with NCC diagnosis by echocardiography and magnetic resonance imaging (21 men, 42.8±14.8 years) were included. A comprehensive echocardiogram was performed. The layer-specific strain was analyzed from the apical two-, three, four-chamber views, short axis, and focused right ventricle views using 2D echocardiography (2DE) software. RBR was present in 44.9% of patients, and this group presented increased LV mass indexed (118±43.4 vs. 94.1±27.1g/m2, P = 0.034), LV end-diastolic and end-systolic volumes (P< 0.001), E/e' (12.2±8.68 vs. 7.69±3.13, P = 0.034), and decreased LV ejection fraction (40.7±8.71 vs. 58.9±8.76%, P < 0.001) when compared to patients without RBR. Also, patients with RBR presented a significant decrease of global longitudinal, radial, and circumferential strain. When ML model based on a random forest algorithm and a neural network model was applied, it found that twist, NC/C, torsion, LV ejection fraction, and diastolic dysfunction are the strongest predictors to RBR with accuracy, sensitivity, specificity, area under the curve of 0.93, 0.99, 0.80, and 0.88, respectively.
CONCLUSION: In this study, a random forest algorithm was capable of selecting the best echocardiographic predictors to RBR pattern in NCC patients, which was consistent with worse systolic, diastolic, and myocardium deformation indices. Prospective studies are warranted to evaluate the role of this tool for NCC risk stratification.

Introduction

Noncompaction cardiomyopathy (NCC) is a genetic cardiomyopathy characterized by prominent left ventricular (LV) trabeculations, deep intertrabecular recesses communicating with the ventricular cavity, and a thin compacted external myocardial layer. Currently, there is no gold standard for the NCC diagnosis, and much debate and discussion persist regarding its classification, pathophysiology, and treatment of this entity [1].

The first echocardiographic description of NCC was reported in 1984, and the first echocardiographic diagnostic criterion was established in 1990 [2]. However, some healthy individuals can fulfill one or more echocardiographic criteria of NCC, and it is not clear if NCC is a distinct cardiomyopathy or an epiphenomenon of other cardiomyopathies [3]. The differential diagnosis with other cardiomyopathies is frequently challenging, and multimodality imaging correlations are usually required, leading to higher costs and extending time until the correct diagnosis is achieved. Although not considered a diagnostic criteria, van Dalen et al. observed that 88% of the NCC patients showed a loss of myocardial twist deformation of LV, with the rotation at the basal and apical levels predominantly in the same direction, a phenomenon called rigid body rotation (RBR) [4]. Another study by Peter et al. also detected this pattern in 53.3% of NCC patients [5], but the clinical relevance of these findings is still uncertain.

In the current era of precision medicine, machine learning techniques has been related to many potential applications in cardiac imaging [6]. Nevertheless, there are few recent papers that explored artificial intelligence tools in NCC patients, above all focused on refining the current diagnostic criteria [7] or on the prediction of adverse clinical outcomes [8] using data extracted from echocardiography and cardiac magnetic resonance, without correlations with myocardial strain analysis. Moreover, circumferential strain analysis software is frequently unavailable in most of echocardiographic machines and this evaluation must be performed offline in dedicated workstations, hampering its availability in small centers or beyond research field.

Although the role of myocardial strain analysis in risk stratification of patients with LV hypertrabeculation remain unknow [9], we hypothesize that RBR may have a potential value in the risk stratification of patients who fulfilled the clinical and imaging criteria of NCC. In that context, we developed and tested ML framework, consisting of a random forest algorithm and a neural network, to evaluate the best echocardiographic predictors to RBR pattern in NCC patients, providing additional information for risk stratification of these patients, without the necessity of image post processing in dedicated workstations.

Materials and methods

This cross-sectional observational study was performed in a quaternary cardiology center where 50 outpatients with a high clinical pre-test probability diagnosis of NCC from 2016 to 2017 were studied [10]. All clinical and echocardiographic data were prospectively collected. NCC was diagnosed when patients fulfill all echocardiographic [11-13] and Petersen [14] criteria by cardiac magnetic resonance imaging, reinforcing the accuracy of this diagnosis in this sample.

Exclusion criteria included pregnancy, valvular heart disease (at least moderate), congenital heart disease, other associated cardiomyopathies, known coronary heart disease, pacemaker, and atrial fibrillation. According to Framingham score or angina, patients over age 40, if they had a moderate risk of coronary artery disease, underwent a non-invasive or invasive coronary artery study. All patients had a complete neurological examination, and all first relatives’ members of patients were also recruited for NCC screening.

Data were expressed as mean ± SD and frequency (percentage). Comparisons between patients with and without RBR were performed using 2-sample t-test (or Wilcoxon rank-sum test) and χ2 tests (or Fisher exact test) for continuous and categorical data, respectively. All statistical analyses were performed using R version 4.0.0 (R Foundation for Statistical Computing, Vienna, Austria), and a P-value of < 0.05 was considered statistically significant. To calculate the correlation matrix, the Pearson form was used, in which the correlation coefficient r for each pair of features and are calculated.

The institutional review board (Comissão de Ética para Análise de Projeto de Pesquisa–CAPPesq, number 0103/09) approved this study for human subject studies, and all participants provided written informed consent before enrollment.

Echocardiographic imaging protocol

Comprehensive echo studies were performed with Vivid E9 echocardiographic ultrasound system (GE Healthcare, Norway). The exams were analyzed by a single trained sonographer, following the European Association of Echocardiography/American Society of Echocardiography guidelines for cardiac chambers and functional analysis [15]. Jenni criterion was used to calculate the N/C ratio, which measures the maximal end-systolic ratio of noncompacted to compacted layers [12]. Data were exported to a dedicated workstation (EchoPac 202, GE Vingmed, Horten, Norway) for off-line analysis by a blinded observer.

Bidimensional left ventricular layer-specific strain

Quantification of layer-specific strain measurements was performed offline with dedicated software (EchoPAC V.202, GE). For speckle tracking echocardiography (STE) and longitudinal strain analysis, digital loops of the right ventricle were obtained from apical 4-chamber and/or right ventricle-focused apical 4-chamber views, and left ventricle (LV) from apical 3-chamber, 2-chamber, and 4-chamber views. For radial and circumferential LV strain, parasternal short axis acquisitions were obtained from the mitral valve, papillary muscle, and apical levels, using mitral valve and apical values to obtain twist and torsion. Three cardiac cycles were acquired from each view at a frame rate of 40–80 frames/sec in patients in sinus rhythm.

Preliminary longitudinal strain analysis was performed online in the ultrasound machine, checking if image quality was good enough to permit adequate tracking of myocardial acoustic markers (speckles) during the entire cardiac cycle. STE analysis was performed semi-automatically, after the operator’s manual setting of 3 points on the endocardial border (2 basal and 1 at the apex). When the region of interest (ROI) included the whole thickness of the ventricle and excluded other structures (such as trabeculae, moderator band, and valvular tissue), the processing was started, and analysis proceeded on a frame-to-frame basis using an automatic tracking system. The ROI generated by the software included basal, mid, and apical segments of opposed walls, divided in 6 segments. Longitudinal peak strain values were measured for each segment, and global longitudinal strain values were calculated by averaging the values. The investigators were blinded to these initial results until the offline analysis of the remaining parameters was performed. Radial and circumferential strains were analyzed exclusively offline. For radial and circumferential strain analysis, the endocardial border was traced just within the endocardium using point-by-point tracing, and particular care was taken to adjust the tracking of all segments. A second larger concentric circle was then automatically generated and manually adjusted near the epicardium such that the area of interest included the entire myocardial wall. The image was then played so that tracking in the region of interest could be fine-tuned by visual assessment to ensure that all wall segments were tracked appropriately throughout the cardiac cycle and that the sectors defining each wall segment were adjusted properly.

Global radial and circumferential strains were measured as the average of the 6 regional segments from the parasternal short-axis view at the mitral valve level, papillary muscles, and LV apex. Left ventricular twist was calculated as the relative rotation of the apex around the LV long axis with respect to the base during the cardiac cycle, and torsion as the twist value normalized to the distance between LV apex and base, expressed in degrees per centimeter (°/cm). LV torsion was considered normal whenever LV presented with initial counterclockwise basal and clockwise apical rotation, followed by end-systolic clockwise basal and counterclockwise apical rotation, and RBR when LV showed rotation at the basal and apical level predominantly in the same direction as demonstrated in Fig 1.

Fig 1

Apical rotation and basal rotation curve obtained from NCC patients.

Normal left ventricular torsion (A), and rigid body rotation pattern (B). The green line represents the apex rotation, and the basal rotation is showed in pink, while the curves mean represented in white. Ordinate axis with positive value expresses counterclockwise rotation and with a negative value, clockwise rotation. AVC, aortic valve closure; LV twist = apical rotation–basal rotation (white line).

Data processing

Fig 2 demonstrates the structure of the proposed ML framework and the main stages involved in its implementation. In this section, these stages will be explained in complete detail. We include the following echocardiogram parameters in this model: twist, torsion, left ventricular ejection fraction, left ventricular global longitudinal strain, left ventricular circumferential strain (apical, mid and basal), left ventricular global circumferential strain, left ventricular radial strain (apical, mid and basal), left ventricular global radial strain, TAPSE, e’, left atrial strain (reservoir, conduit and buster), left ventricular eccentric and concentric hypertrophy, tricuspid velocity peak, isovolumic relaxation time, right ventricular free wall strain, right ventricular S’, relative wall thickness, A wave velocity, E wave velocity, E/A ratio, E/e’, diastolic dysfunction left ventricular length, deceleration time, fractional area change, left atrial volume index, right ventricle (base region), left ventricular diastolic and systolic diameters, left ventricular end systolic and end diastolic index volume, NC/C ratio, RBR pattern.

Fig 2

Task execution diagram from the data acquisition to the results of noncompaction cardiomyopathy patients.

Initially, the raw data samples were pre-processed to provide a set of input samples that can be used for developing an efficient learning model. For this purpose, features (columns) or samples (lines) that had many null values or values outside the range were eliminated from the bank. Both situations, removing columns or rows, were carefully analyzed because removing a critical column can harm the model’s performance, whereas removing rows can reduce the size of the bank, making the training and testing stage unfeasible.

The number of features was also reduced by feature engineering techniques that allow the choice of the best features based on their influence on the performance of a learning model. A feature selection method based on a random forest algorithm was used [16]. This method indicates how useful a feature is for building the classifier. A tree was created for each feature, and then the number of optimization divisions for each tree is observed. The performance measure consists of the error function of the tree by the number of divisions.

Multilayer perceptron neural networks

A neural network model called Multi-Layer Perceptron (MLP) was used. This type of neural network was selected to solve complicated classification problems when the amount of data for training the network is limited. However, the proposed ML framework can also be implemented using a state-of-the-art deep neural network when an extensive dataset for properly training such a network is available.

For adequate configuration of a neural network, layers, and neurons, it is necessary to consider factors such as the quality and quantity of samples from the dataset and the nature of the addressed problem. However, a close configuration is possible based on an empirical approach and observing criteria like regular distribution of layers and error rate. A small number of hidden layers can generate an inaccurate model, whereas too many layers can cause overfitting. Also in Fig 2, the input layer contains the number of neurons corresponding to the features; the hidden layers are distributed in 3 layers, with the same activation function, rectified linear unit (ReLU); however, the output layer counts on 2 neurons and uses a sigmoidal activation function.

The Scikit-learn library was used to provide the implementation of this neural network. A k-fold validation approach with k = 5 was used to validate the model, corresponding to a training dataset with 60% of the samples, testing with 20%, and validation with 20%. The training stages of 3000 times each, and the learning rate is equal to 0.001.

Results

Table 1 summarizes the demographic data in the present study. A total of 50 NCC patients were recruited, and one patient was excluded because more than 2 segments of a 17-segment model had low quality for strain analysis. Table 1 summarizes the demographic data in the entire sample (n = 49) and in the patients with (n = 22) or without (n = 27) LV RBR. Twenty-one men (42.8 ± 14.8 years) were included in the sample. The most common cardiovascular risk factor was systemic arterial hypertension (26.5%); most patients were in NYHA I (79.6%); family history of NCC was present in almost half of patients (49%). Anticoagulation therapy was statistically significant in the left ventricular rigid body rotation (LV-RBR) group (90.9% vs. 70.4%, P = 0.019).

Table 1

Demographic, clinical, and laboratory data of noncompaction patients.

Patient characteristics		All patients (N = 49)	LV-RBR absent (n = 27)	LV-RBR present (n = 22)	P -value
Gender (male), n (%)		21 (42.9)	10 (37.0)	11 (50.0)	0.534
Age (years)		42.8 (14.8)	43.8 (12.6)	41.5 (17.3)	0.596
Body surface area (m2)		1.76 (0.21)	1.81 (0.17)	1.69 (0.24)	0.068
NYHA, n (%)	I	39 (79.6)	22 (81.5)	17 (77.3)	0.868
	II	7 (14.3)	3 (11.1)	4 (18.2)	0.238
	III	3 (6.12)	2 (7.41)	1 (4.55)	0.037
Diabetes mellitus, n (%)		3 (6.12)	3 (11.1)	0 (0.00)	0.242
Smokers, n (%)		3 (6.12)	2 (7.41)	1 (4.55)	1.000
Systemic arterial hypertension, n (%)		13 (26.5)	8 (29.6)	5 (22.7)	0.827
Ventricular tachycardia, n (%)		7 (14.3)	4 (14.8)	3 (13.6)	1.000
Family history of NCC, n (%)		24 (49.0)	15 (55.6)	9 (40.9)	0.464
History of embolic events, n (%)		1 (2.04)	0 (0.00)	1 (4.55)	0.449
ACE-inhibitor, n (%)		42 (85.7)	22 (81.5)	20 (90.9)	0.436
Anticoagulation, n (%)		19 (38.8)	6 (22.2)	13 (59.1)	0.019
Beta blocking agent, n(%)		39 (79.6)	19 (70.4)	20 (90.9)	0.152
Serum creatinine (mg/dL)		0.90 (0.20)	0.91 (0.21)	0.88 (0.19)	0.593

ACE: angiotensin-converting enzyme; LV-RBR: left ventricular rigid body rotation; NCC: noncompaction cardiomyopathy; NYHA: New York Heart Association (functional class).

A total of 38 echocardiographic features were extracted from the dataset, as described in Table 2. Rigid body rotation pattern was present in 44.9% of patients. Left ventricular remodeling assessed by LV mass indexed (118±43.4 vs. 94.1±27.1g/m2, P = 0.034), left ventricular diastolic diameter (55.4±7.45 vs. 49.3±6.20mm, P = 0.004) and the LV end-diastolic volume index (92.7±38.9 vs. 53.6±14.1mL/m2, P < 0.001) were higher in the RBR group. Also, E/e’ was increased (12.2±8.68 vs. 7.69±3.13, P = 0.034) and LV ejection fraction (40.7±8.71 vs. 58.9±8.76%, P < 0.001) was lower in patients with RBR. Left ventricular mechanical parameters such as global longitudinal, radial, and circumferential strain were more affected in the RBR group, whereas noncompaction/compaction ratio (NC/C) presented no difference. On the other hand, LV diastolic parameter, left atrial strain, and right ventricle free wall strain (RV FWS) were not statistically different between the groups (Table 2).

Table 2

Echocardiographic parameters evaluated in noncompaction cardiomyopathy with rigid body rotation and included in the prediction model.

Echocardiographic features, n (%)		All patients (N = 49)	LV-RBR absent (n = 27)	LV-RBR present (n = 22)	P-value
LV mass (g)		183 (63.3)	171 (54.1)	198 (71.6)	0.152
LVIM (g/m2)		105 (36.9)	94.1 (27.1)	118 (43.4)	0.034
LVDD (mm)		52.0 (7.39)	49.3 (6.20)	55.4 (7.45)	0.004
RV basal (mm)		34.9 (6.16)	34.0 (4.83)	36.0 (7.44)	0.265
TAPSE (mm)		21.7 (4.35)	23.2 (3.53)	19.7 (4.55)	0.005
S’ RV (cm/s)		13.2 (2.64)	13.8 (2.06)	12.3 (3.10)	0.068
FAC (%)		49.2 (10.8)	47.3 (8.56)	51.4 (12.8)	0.202
LAI (mL/m2)		35.2 (14.8)	32.6 (14.1)	38.4 (15.4)	0.184
E (cm/s)		76.4 (23.9)	73.1 (21.3)	80.5 (26.8)	0.314
A (cm/s)		59.1 (19.7)	57.5 (17.1)	61.2 (22.8)	0.532
E/A		1.37 (0.44)	1.32 (0.36)	1.43 (0.53)	0.416
DT (ms)		219 (85.8)	202 (78.3)	239 (92.3)	0.155
TRIV (ms)		111 (25.3)	109 (28.6)	113 (21.0)	0.640
e’ (cm/s)		9.78 (4.40)	10.7 (4.43)	8.60 (4.16)	0.097
E/e’		9.73 (6.58)	7.69 (3.13)	12.2 (8.68)	0.031
Tricuspid velocity peak (m/s)		3.26 (4.67)	2.46 (0.68)	4.02 (6.47)	0.354
LV EDVI (mL/m2)		71.1 (34.0)	53.6 (14.1)	92.7 (38.9)	<0.001
LV ESVI (mL/m2)		38.1 (28.0)	22.5 (9.71)	57.2 (31.3)	<0.001
LVEF (%)		50.7 (12.6)	58.9 (8.76)	40.7 (8.71)	<0.001
LV GLS (%)		15.4 (4.67)	17.9 (3.74)	12.2 (3.72)	<0.001
LV Twist (°)		9.40 (8.32)	15.2 (6.25)	2.32 (3.82)	<0.001
LV Torsion (°/cm)		1.16 (1.05)	1.89 (0.81)	0.27 (0.45)	<0.001
LAS reservoir (%)		30.6 (11.5)	32.3 (10.8)	28.5 (12.2)	0.256
LAS booster (%)		12.1 (6.09)	12.6 (6.05)	11.5 (6.22)	0.517
LAS conduit (%)RV FWS (%)		18.1 (9.40)	19.6 (9.22)	16.2 (9.50)	0.213
		23.0 (7.50)	24.1 (5.55)	21.6 (9.30)	0.265
LVRS basal (%)		26.3 (14.9)	33.2 (14.0)	17.9 (11.6)	<0.001
LVRS mid (%)		26.9 (15.5)	30.0 (15.4)	23.1 (15.1)	0.121
LVRS apical (%)		22.8 (18.5)	23.7 (13.7)	21.7 (23.4)	0.720
LV GRS (%)		25.3 (11.5)	29.0 (10.8)	20.9 (11.1)	0.014
LVCS basal (%)		11.0 (6.32)	13.2 (4.95)	8.16 (6.79)	0.006
LVCS mid (%)		11.5 (6.18)	13.2 (4.76)	9.53 (7.18)	0.049
LVCS apical (%)		13.7 (7.69)	17.0 (6.23)	9.57 (7.44)	0.001
LV GCS (%)		12.1 (6.16)	14.5 (4.37)	9.22 (6.88)	0.004
Dilated LV		0.22 (0.42)	0.11 (0.32)	0.36 (0.49)	0.046
NC/C		2.99 (±0.99)	2.3 (±0.15)	3.72 (±0.98)	0.13
LV (%) remodeling	CH	6 (12.2)	4 (14.8)	2 (9.09)	0.940
	CR	4 (8.16)	2 (7.41)	2 (9.09)	0.572
	EH	16 (32.7)	8 (29.6)	8 (36.4)	0.940
	normal	23 (46.9)	13 (48.1)	10 (45.5)	0.957

CH, concentric hypertrophy; CR, concentric remodeling; DT, deceleration time; EH, eccentric hypertrophy; FAC, fractional area change; IRVT, isovolumic relaxation time; LAI, left atrium volume indexed; LAS, left atrium strain; LV, left ventricle; LVDD, left ventricular diastolic diameter; LV EDVI, left ventricular end-diastolic volume indexed; LVEF, left ventricle ejection fraction; LVCS, left ventricular circumferential strain; LV ESVI, left ventricular end-systolic volume indexed; LV GCS, left ventricular global circumferential strain; LV GLS, left ventricular global longitudinal strain; LV GRS, left ventricular global radial strain; LVIM, left ventricular mass indexed; LVRS, left ventricular radial strain; NC/C, noncompacted/compacted ratio; RBR, left ventricular rigid body rotation; RV, right ventricle; RV FWS, right ventricle free wall strain; TAPSE, tricuspid annular plane systolic excursion.

The correlation matrix of the used dataset is shown in the Fig 3; it is possible to identify a set of features that have a high correlation with RBR. The strongest ones were torsion, twist (r = -0.65), NC/C (r = 0.62), and LVEF (r = -0.51). Additional findings, the ratio NC/C had the following correlations: LVEF (r = - 0.77), LV diastolic dysfunction (r = 0.63), LV GLS (r = -0.62), torsion (r = -0.61), and twist (r = -0.60).

Fig 3

Pearson’s correlation matrix is ordered from the coefficients between the features and rigid body rotation.

Fig 4 presents the features in decreasing order, according to the influence on the Random Forest classification. The features marked as blue were selected based on the feature selection algorithm, and the feature marked as green was selected based on medical decision.

Fig 4

Echocardiographic features in decreasing order of importance to rigid body rotation in noncompaction cardiomyopathy patients.

Ten features were selected based on the results obtained after applying the importance ranking and recursive elimination methods. One additional feature LV GLS was included in the model due to its clinical relevance. This type of feature plays an essential role during the classification. However, the algorithm can ignore this importance due to factors like small samples and slight variation in the value of the feature in the samples, justifying the correction by medical analysis in specific occasion. Finally, the features marked in red were removed from the model formation after the application of feature selection and medical analysis.

Aiming to select which parameters influence the RBR status in NCC patients, our model achieved high accuracy, sensitivity, and specificity: 0.93, 0.99, and 0.88 using the dataset with features selection approach. The area under the curve (AUC) in Fig 5 illustrates the performance of our model, which was 0.92.

Fig 5

ROC curves for prediction model to rigid body rotation pattern in noncompaction cardiomyopathy patients.

The continuous blue line shows the ROC curve obtained from the model’s performance during the training stage. The dashed orange line shows the ROC curve obtained during the test step.

The classification was also performed using the database with all the features. This execution aimed to identify whether the application of the features selection technique had positive results in the model’s performance. The results obtained for accuracy, sensitivity, and specificity: 0.86, 0.88, and 0.83, respectively.

Discussion

To the best of our knowledge, this is the first study assessing RBR by LV mechanics in NCC patients using a neural network to understand the intricate interaction between different echocardiographic parameters. For this, we have analyzed which echocardiographic parameters can consistently predict the presence of RBR, providing a new approach in the pathophysiological mechanism of LV contraction in NCC patients.

Left ventricular torsion occurs as a balance between the interaction of endocardial and epicardial fibers [17]. It is acceptable that the disappearance of torsion could increase endocardial stress and strain, increasing heart oxygen demand [18]. Conversely, NCC may present a distinct pattern [4], characterized by clockwise basal and apical rotation throughout systole. The excess of trabeculation affects the endocardial layer. Thereby, the more trabeculated is the LV, the more severe endocardial fibers are affected. It could be an explanation for why RBR pattern is more prevalent in NCC patients.

Recently, Sabatino et al. proposed a discriminative value of LV twist in the NCC diagnosis [19]. Despite that, twist analysis has limited availability because it is laborious and not incorporated into most echocardiographic machines. Also, images acquisition/storage in a workstation offline is needed. Computational models could automatize these measurements and the algorithms are sufficiently close to automated measurements, reducing the time spent on analysis and errors related to manual calculations [20], especially in NCC patients.

Much debate persists regarding the clinical relevance of RBR pattern in NCC patients. Peters et al. showed that RBR was not associated with more adverse remodeling than in subjects with LVNC and normal LV rotation [5]. However, different from our study, they showed that LVEF was severely reduced in both groups (LVEF 27.9 ± 9.7 vs. 24.9 ± 11.7), making it more difficult to find consistent differences in LV remodeling. Of note, the RBR pattern was found in 53.3% of patients, remarkably close to our findings (44.9%). On the other hand, we found that RBR pattern was associated with worse LV remodeling and lower LVEF, increased ventricular volumes and mass, which is more consistent with previous studies in other cardiomyopathies [21, 22]. Interestingly, our matrix correlation and AI model recognized that RBR was correlated with 5 relevant echocardiographic parameters: LVEF, diastolic dysfunction twist, NC/C ratio and torsion.

Nowadays, which mechanical parameters have prognostic relevance in NCC patients are considered an open question. Intriguingly, the Multi-Ethnic Study of Atherosclerosis (MESA) study, designed to investigate the prevalence and progression of subclinical atherosclerosis (not including NCC patients), found that LV function was worse in individuals with greater LV trabeculation [9]. Gastl et al. showed that cardiac magnetic resonance imaging derived deformation indices may show added value to assess functional impairment in LVNC, regardless of LVEF [23]. We found that the NC/C ratio correlates with RBR, LVEF, and diastolic dysfunction, which adds a coherent link between excess trabeculation and LV mechanical disturbance. Prospective studies are warranted to further explore these correlations in order to find new diagnostic and prognostic markers in NCC patients.

Our findings might suggest new insights into progressive myocardial dysfunction and the pathogenesis of NCC. We emphasize that these tools are not ready for clinical use, but we think they may impact the clinical decision-making in the near future. We also recognize that our results should be interpreted as an exploratory analysis, and they are not a substitute for the complete analysis of myocardial strain when available. Prospective studies with larger patient cohorts are needed to further validate the prognostic implication of our data, and it lends support for the role of this technique in refining the risk stratification of NCC patients.

Limitations

This study has some limitations. First, a small sample size limits the application of more advanced neural networks. Second, we could not explore quantitatively the parameters that caused more influence in RBR, considering our small sample size. Third, all analysis was performed in only one software, we could not test the algorithm with different vendors which reduce its applicability. While there is a joint effort among the vendors to diminish these discrepancies [24]. The database used has a challenging configuration when working with machine learning techniques such as neural networks due to the number of samples and features [25]. Nevertheless, it was possible to use feature engineering techniques that make it possible to adjust the database for the desired application, maintaining the consistency of the results.

Conclusion

Rigid body rotation was associated with pronounced LV remodeling and dysfunction. A machine learning model could identify the 11 parameters that consistently predict the presence of RBR in NCC patients. Further prospective studies should be addressed in order to investigate the role of RBR in the diagnosis and prognostication of NCC patients.

16 Jun 2021

PONE-D-21-10289

A Machine learning framework for the evaluation of myocardial dysfunction in patients with noncompaction cardiomyopathy and rigid body rotation

PLOS ONE

Dear Dr. Salemi,

Thank you for submitting your manuscript to PLOS ONE. After careful consideration, we feel that it has merit but does not fully meet PLOS ONE’s publication criteria as it currently stands. Therefore, we invite you to submit a revised version of the manuscript that addresses the points raised during the review process.

Please submit your revised manuscript by Jul 31 2021 11:59PM. If you will need more time than this to complete your revisions, please reply to this message or contact the journal office at plosone@plos.org. When you're ready to submit your revision, log on to https://www.editorialmanager.com/pone/ and select the 'Submissions Needing Revision' folder to locate your manuscript file.

Please include the following items when submitting your revised manuscript:

A rebuttal letter that responds to each point raised by the academic editor and reviewer(s). You should upload this letter as a separate file labeled 'Response to Reviewers'.

A marked-up copy of your manuscript that highlights changes made to the original version. You should upload this as a separate file labeled 'Revised Manuscript with Track Changes'.

An unmarked version of your revised paper without tracked changes. You should upload this as a separate file labeled 'Manuscript'.

If you would like to make changes to your financial disclosure, please include your updated statement in your cover letter. Guidelines for resubmitting your figure files are available below the reviewer comments at the end of this letter.

If applicable, we recommend that you deposit your laboratory protocols in protocols.io to enhance the reproducibility of your results. Protocols.io assigns your protocol its own identifier (DOI) so that it can be cited independently in the future. For instructions see: http://journals.plos.org/plosone/s/submission-guidelines#loc-laboratory-protocols. Additionally, PLOS ONE offers an option for publishing peer-reviewed Lab Protocol articles, which describe protocols hosted on protocols.io. Read more information on sharing protocols at https://plos.org/protocols?utm_medium=editorial-email&utm_source=authorletters&utm_campaign=protocols.

We look forward to receiving your revised manuscript.

Kind regards,

Zhifan Gao

Academic Editor

PLOS ONE

Journal Requirements:

When submitting your revision, we need you to address these additional requirements.

1. Please ensure that your manuscript meets PLOS ONE's style requirements, including those for file naming. The PLOS ONE style templates can be found at and https://journals.plos.org/plosone/s/file?id=ba62/PLOSOne_formatting_sample_title_authors_affiliations.pdf

2. Thank you for including your ethics statement: This study was approved by the institutional review board for human subject studies, and all participants provided written informed consent prior to enrollment.

a) Please amend your current ethics statement to include the full name of the ethics committee/institutional review board(s) that approved your specific study.

Once you have amended this/these statement(s) in the Methods section of the manuscript, please add the same text to the “Ethics Statement” field of the submission form (via “Edit Submission”).

For additional information about PLOS ONE ethical requirements for human subjects research, please refer to http://journals.plos.org/plosone/s/submission-guidelines#loc-human-subjects-research.

3.  Thank you for stating the following financial disclosure:

[No].

At this time, please address the following queries:

Please clarify the sources of funding (financial or material support) for your study. List the grants or organizations that supported your study, including funding received from your institution.

State what role the funders took in the study. If the funders had no role in your study, please state: “The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.”

If any authors received a salary from any of your funders, please state which authors and which funders.

If you did not receive any funding for this study, please state: “The authors received no specific funding for this work.”

Please include your amended statements within your cover letter; we will change the online submission form on your behalf.

4.  Thank you for stating the following in your Competing Interests section:

[No].

Please complete your Competing Interests on the online submission form to state any Competing Interests. If you have no competing interests, please state "The authors have declared that no competing interests exist.", as detailed online in our guide for authors at http://journals.plos.org/plosone/s/submit-now

This information should be included in your cover letter; we will change the online submission form on your behalf.

Please know it is PLOS ONE policy for corresponding authors to declare, on behalf of all authors, all potential competing interests for the purposes of transparency. PLOS defines a competing interest as anything that interferes with, or could reasonably be perceived as interfering with, the full and objective presentation, peer review, editorial decision-making, or publication of research or non-research articles submitted to one of the journals. Competing interests can be financial or non-financial, professional, or personal. Competing interests can arise in relationship to an organization or another person. Please follow this link to our website for more details on competing interests: http://journals.plos.org/plosone/s/competing-interests

Additional Editor Comments (if provided):

[Note: HTML markup is below. Please do not edit.]

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Partly

Reviewer #2: Partly

**********

2. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: Yes

**********

3. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: No

Reviewer #2: No

**********

4. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

**********

5. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: In this study, authors presented a study of NCC identification by using ML method. They focused on evaluation of the RBR criteria. They stated that good performance was obtained. However, there were several major issues that the result was therefore not convincing.

Comments:

1. The language usage throughout this paper need to be improved, the author should do some proofreading on it. Give the article a mild language revision to get rid of few complex sentences that hinder readability and eradicate typo erros.

2. Overall, the basic background is not introduced well. To my understanding, differentiation between NCC patient to non NCC patient was the key objective. However, no explanation about the ‘3 most common criteria’ was provided while limited information about the RBR without explanation about the correlation between RBR and NCC and the correlation between RBR and other echocardiographic parameters. It was confusing that whether the NCC or the RBR was the objective of this work.

3. Introduction needs to explain the main contributions of the work more clearly. The novelty of this paper is not clear.

4. Literature review techniques has to be strengthened. The authors should consider more recent research done in the field of their study.

5. The method section is not clear. Above all, it is not clear that what kinds of parameters were collected. How did you collected the parameters that were not mentioned in method section. What is the characteristic of the data. Second, it is not clear whether the image quality check was only performed on data for LS analysis? What about the other parameters?

6. There is a missing of citation for the random forest algorithm. How did you determine the key features? It is not clear about the architecture of the MLP. Also, it is not clear about the split of data for training, validation and test.

7. Result section. Correlation was presented, but where was the result of RF for parameters selection?

8. It is unclear how NC/C was generated.

9. Discussion section. what is boundary subsets?

10. Authors stated that ‘cardiac mechanics seems to be more promising as a specific marker, and trabeculae may be considered just a red flag cardiac mechanics’. However, while the cardiac mechanics parameters, ie. GLS and LVEF were the result of various phenotypes, it is not clear how the authors came to this statement. Please elaborate.

Reviewer #2: This paper uses a machine learning (ML) based model to evaluate myocardium function, and the role of echocardiographic predictors for rigid body rotation in Noncompaction cardiomyopathy (NCC). The manuscript is clear, straightforward, easy to follow. The results are sufficient and convincing. I think this study has the potential for NCC risk stratification. However, I would like to draw the author's attention to the following major concerns:

1)In this study, the authors used 50 patients for training and testing the machine learning model. Considering, the model is a deep neural network. The data set the size of 50 patients is insufficient.

2)The authors do not explain clearly its used deep neural network. As a key point, what is the architecture of the deep neural network? Does the network use a general architecture, or is it specially optimized for NCC? Do different network architectures change the results？The authors should do a more thorough literature survey. Just to name a few:

-Cardiac Functional Analysis with Cine MRI via Deep Learning Reconstruction

-Segmentation and quantification of infarction without contrast agents via spatiotemporal generative adversarial learning.

-Automated left ventricular dimension assessment using artificial intelligence developed and validated by a UK-wide collaborative

3) In this study, the authors used a single echocardiographic ultrasound system from a single center. I think this point is the most important limitation of this study. The generalization error is a big problem for machine learning (ML) based models. It would be better to discuss this point in the manuscript.

4) There are some grammar errors and typos. I suggest the authors make an solid, overall proofreading.

**********

6. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

[NOTE: If reviewer comments were submitted as an attachment file, they will be attached to this email and accessible via the submission site. Please log into your account, locate the manuscript record, and check for the action link "View Attachments". If this link does not appear, there are no attachment files.]

While revising your submission, please upload your figure files to the Preflight Analysis and Conversion Engine (PACE) digital diagnostic tool, https://pacev2.apexcovantage.com/. PACE helps ensure that figures meet PLOS requirements. To use PACE, you must first register as a user. Registration is free. Then, login and navigate to the UPLOAD tab, where you will find detailed instructions on how to use the tool. If you encounter any issues or have any questions when using PACE, please email PLOS at figures@plos.org. Please note that Supporting Information files do not need this step.

12 Oct 2021

Dear Dr. Gao

We are thrilled by the opportunity to resubmit our revised manuscript. On behalf of all co-authors, I would like to thank you and the invited reviewers for the time and effort dedicated to reviewing our work, as well as for the revisions that we believe have strengthened this paper.

We have made every effort to address all points and issues raised. Enclosed please see detailed response to each of the reviewers’ comments to the manuscript. We believe that the information provided in this manuscript has significant clinical value and should be communicated to the scientific community through this journal.

Sincerely,

Submitted filename: Response to reviewrs.docx

Click here for additional data file.

5 Nov 2021

A Machine learning framework for the evaluation of myocardial dysfunction in patients with noncompaction cardiomyopathy and rigid body rotation

PONE-D-21-10289R1

Dear Dr. Salemi,

We’re pleased to inform you that your manuscript has been judged scientifically suitable for publication and will be formally accepted for publication once it meets all outstanding technical requirements.

Within one week, you’ll receive an e-mail detailing the required amendments. When these have been addressed, you’ll receive a formal acceptance letter and your manuscript will be scheduled for publication.

An invoice for payment will follow shortly after the formal acceptance. To ensure an efficient process, please log into Editorial Manager at http://www.editorialmanager.com/pone/, click the 'Update My Information' link at the top of the page, and double check that your user information is up-to-date. If you have any billing related questions, please contact our Author Billing department directly at authorbilling@plos.org.

If your institution or institutions have a press office, please notify them about your upcoming paper to help maximize its impact. If they’ll be preparing press materials, please inform our press team as soon as possible -- no later than 48 hours after receiving the formal acceptance. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information, please contact onepress@plos.org.

Kind regards,

Zhifan Gao

Academic Editor

PLOS ONE

Additional Editor Comments (optional):

Reviewers' comments:

Reviewer's Responses to Questions

Comments to the Author

1. If the authors have adequately addressed your comments raised in a previous round of review and you feel that this manuscript is now acceptable for publication, you may indicate that here to bypass the “Comments to the Author” section, enter your conflict of interest statement in the “Confidential to Editor” section, and submit your "Accept" recommendation.

Reviewer #1: All comments have been addressed

Reviewer #2: All comments have been addressed

**********

2. Is the manuscript technically sound, and do the data support the conclusions?

The manuscript must describe a technically sound piece of scientific research with data that supports the conclusions. Experiments must have been conducted rigorously, with appropriate controls, replication, and sample sizes. The conclusions must be drawn appropriately based on the data presented.

Reviewer #1: Yes

Reviewer #2: Yes

**********

3. Has the statistical analysis been performed appropriately and rigorously?

Reviewer #1: Yes

Reviewer #2: Yes

**********

4. Have the authors made all data underlying the findings in their manuscript fully available?

The PLOS Data policy requires authors to make all data underlying the findings described in their manuscript fully available without restriction, with rare exception (please refer to the Data Availability Statement in the manuscript PDF file). The data should be provided as part of the manuscript or its supporting information, or deposited to a public repository. For example, in addition to summary statistics, the data points behind means, medians and variance measures should be available. If there are restrictions on publicly sharing data—e.g. participant privacy or use of data from a third party—those must be specified.

Reviewer #1: No

Reviewer #2: No

**********

5. Is the manuscript presented in an intelligible fashion and written in standard English?

PLOS ONE does not copyedit accepted manuscripts, so the language in submitted articles must be clear, correct, and unambiguous. Any typographical or grammatical errors should be corrected at revision, so please note any specific errors here.

Reviewer #1: Yes

Reviewer #2: Yes

**********

6. Review Comments to the Author

Please use the space provided to explain your answers to the questions above. You may also include additional comments for the author, including concerns about dual publication, research ethics, or publication ethics. (Please upload your review as an attachment if it exceeds 20,000 characters)

Reviewer #1: Authors presented an exploratory investigation of the markers that could potentially help in differentiate diagnosis for NCC patients. Significant improvements have been made. There are several minor issues

1. In Figure 1, it seems that the white line was supposed to be the difference between green line and pink line, while it is confusing in the caption that what does the “curve mean” mean?

2. It will be appreciated to see the ROC analysis for differentiating the cases with or without RBR in patients with NCC.

Reviewer #2: (No Response)

**********

7. PLOS authors have the option to publish the peer review history of their article (what does this mean?). If published, this will include your full peer review and any attached files.

If you choose “no”, your identity will remain anonymous but your review may still be made public.

Do you want your identity to be public for this peer review? For information about this choice, including consent withdrawal, please see our Privacy Policy.

Reviewer #1: No

Reviewer #2: No

11 Nov 2021

PONE-D-21-10289R1

A machine learning framework for the evaluation of myocardial rotation in patients with noncompaction cardiomyopathy

Dear Dr. Salemi:

I'm pleased to inform you that your manuscript has been deemed suitable for publication in PLOS ONE. Congratulations! Your manuscript is now with our production department.

If your institution or institutions have a press office, please let them know about your upcoming paper now to help maximize its impact. If they'll be preparing press materials, please inform our press team within the next 48 hours. Your manuscript will remain under strict press embargo until 2 pm Eastern Time on the date of publication. For more information please contact onepress@plos.org.

If we can help with anything else, please email us at plosone@plos.org.

Thank you for submitting your work to PLOS ONE and supporting open access.

Kind regards,

PLOS ONE Editorial Office Staff

on behalf of

Dr. Zhifan Gao

Academic Editor

PLOS ONE