Novel Deep Learning Framework for Simultaneous Assessment of Left Ventricular Mass and Longitudinal Strain: Clinical Feasibility and Validation in Patients with Hypertrophic Cardiomyopathy

1 Aims: This study aims to present the Segmentation-based Myocardial Advanced Refinement 2 Tracking (SMART) technique, a novel artificial intelligence (AI) -based approach for 3 transthoracic echocardiography (TTE) that incorporates motion tracking and left ventricular 4 (LV) myocardial segmentation throughout the cardiac cycle. The study evaluates its feasibility 5 and accuracy in assessing LV mass (LVM) and global longitudinal strain ( LVGLS) in 6 hypertrophic cardiomyopathy (HCM) patients. 7

and results: The study included 111 HCM patients (median age: 58 years, 69% male) 8 who underwent TTE and cardiac magnetic resonance imaging (CMR). SMART -derived 9 LVGLS (LVGLS SMART) showed strong correlation with conventional manual LVGLS 10 measurements ( LVGLSTTE-EchoPAC; Pearson’s correlation coefficient [PCC] 0.851; mean 11 difference 0 [-2–0]). When compared to CMR as the reference standard for LVM measurement 12 (LVMCMR), conventional dimension-based TTE calculations (LVMTTE-DB) overestimated LVM 13 compared to CMR ( PCC 0.652; mean difference: 106 [90 –123]), while LVM SMART 14 demonstrated excellent agreement with CMR (PCC 0.843; mean difference: 1 [-11–13]). For 15 predicting extensive LGE (≥15% of LVM), LVGLSSMART and LVMSMART exhibited comparable 16 predictive performance (AUC: 0.72 and 0.66, respectively) to conventional LVGLS and CMR. 17 Patients identified as high -risk for extensive LGE by LVGLS SMART (≤11%) and LVM SMART 18 (≥166g) had significantly higher rates of adverse outcomes, including heart failure 19 hospitalization, new-onset atrial fibrillation, and defibrillator implantation. 20

The SMART technique provides a comparable evaluation of LVGLS and a more 21 accurate assessment of LVM compared to conventional TTE, with additional predictive value 22 for myocardial fibrosis and adverse outcomes, supporting its utility in HCM management. 23 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 5

cardiac magnetic resonance imaging; echocardiography; global longitudinal strain; 1 hypertrophic cardiomyopathy; left ventricle; mass. 2 3 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 6 1. INTRODUCTION 1 Transthoracic echocardiography (TTE) is the primary diagnostic tool for hypertrophic 2 cardiomyopathy (HCM).(1) Evaluation of HCM typically involves assessing left ventricular 3 (LV) hypertrophy and function through visual estimation and quantitative measurements, such 4 as LV wall dimensions and ejection fraction (EF). (2) However, this conventional approach is 5 operator-dependent, time -consuming, and often incomplete. In particular, the thickened LV 6 wall in HCM reduces cavity size, leading to overestimation of EF, making LV global 7 longitudinal strain (LVGLS) a more accurate measure of myocardial contractility. Furthermore, 8 dimension-based LV mass (LVM) calculation fail s to accurately reflect asymmetric 9 characteristics of LV hypertrophy in HCM. (3, 4) Consequently, cardiac magnetic resonance 10 (CMR) is often required for a more precise evaluation.(1) 11 Recent advances in artificial intelligence (AI) have enabled automated segmentation 12 of LV cavity and wall, providing LV volume, EF, and mass in TTE. (5-7) Automated LVGLS 13 measurement is also now available and validated across various populations. (8, 9) However, 14 the unique geometry of HCM complicates the automated detection of LV endocardial and 15 epicardial borders, as well as accurate motion tracking. Consequently, validation of AI -based 16 automated measurements remains limited in HCM, particularly in accurately measuring 17 LVLVM in cases with asymmetric hypertrophy. 18 Our research group has previously developed and validated a comprehensive 19 automated system for TTE analysis, (10-15) featuring an algorithm that simultaneously 20 segments the LV wall and track s myocardial motion, validated in myocardial infarction 21 patients.(14) Building on this, we developed the Segmentation -based Myocardial Advanced 22 Refinement Tracking (SMART) technique, which refines motion tracking using LV myocardial 23 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 7 segmentation throughout the cardiac cycle. In this study, we aimed to present the SMART 1 technique and evaluate its ability to successfully segment and track the LV wall in HCM 2 patients, enabling accurate assessment of LVM and LVGLS. 3 4 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 8 2. METHODS 1 2.1. Study Population 2 The AI-based system utilized in this study was developed and validated with data from the 3 Open AI Dataset Project (AI -Hub), which was initiated by the Ministry of Science and ICT, 4 Korea.(10-16) For the development of the SMART technique, no additional developmental 5 datasets were required beyond those used in the original automatic segmentation and strain 6 analysis.(14) (Supplemental Method 1 ) We conducted in -house validation of the SMART 7 technique across diverse study populations , including the internal test set, an American 8 population, and a paired dataset with both TTE and CMR. (Supplemental Method 1) 9 In this study, t o further evaluate the feasibility and accuracy of our novel SMART 10 technique in HCM, we used data from a previous study cohort of 111 patients with HCM who 11 underwent both TTE and CMR within a 6 -month at Seoul National University Bundang 12 Hospital between 2010 and 2019. (17) The clinical diagnosis of HCM was established by 13 identifying a maximal end -diastolic wall thickness >15 mm in any segment of the LV , with 14 other potential causes of hypertrophy excluded. The study protocol was approved by the 15 Institutional Review Board of Seoul National University Bundang Hospital (IRB No. B-2305-16 827-002), with a waiver of informed consent granted due to the retrospective study design. All 17 clinical data were fully anonymized prior to analysis. The study was conducted in accordance 18 with the principles outlined in the Declaration of Helsinki (2013). 19 20 2.2. TTE Acquisition and Analysis 21 All echocardiographic examinations were performed by trained echocardiographers or 22 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 9 cardiologists and interpreted by board -certified cardiologists with expertise in 1 echocardiography, following current guidelines. Standard ultrasound machines (Vivid 7, E9, 2 and E95; GE Vingmed Ultrasound AS, Norway) were used with a 2.5 -MHz probe at 50 to 80 3 Hz frame rates . Following standard protocols, 2 -dimensional (2D), M -mode, and Doppler 4 images were obtained. 5 LV maximal wall thickness (LVMWTTTE) was measured at the thickest segment of the 6 LV wall in end-diastole (ED). Dimension-based LVM (LVMTTE-DB) was calculated using the 7 interventricular septum (IVS), left ventricular internal diameter (LVID), and posterior wall 8 thickness (PWT) in ED according to the Devereux equation (3): 9 0.8 × 1.04 × [(IVS + LVID + PWT)3 − LVID3] + 0.6 g 10 LV ED and end -systolic (ES) volumes (LVEDV TTE, and LVESV TTE, respectively) were 11 measured from the apical 4-chamber view (A4C) and apical 2-chamber view (A2C) to calculate 12 LVEF using the Simpson biplane method. (2) Tissue Doppler imaging was used to assess the 13 early mitral annular velocity (e’) at the septal side of the mitral annulus. 14 LVGLS analysis was performed by an imaging specialist, using two different software 15 programs: TomTec Image Arena 4.6 (Munich, Germany) for endocardial tracking (LVGLSTTE-16 Tomtec) and EchoPAC PC BT20 (GE Medical Systems, Norway) for mid -myocardial tracking 17 (LVGLSTTE-EchoPAC). Both analyses utilized the same TTE image frames focusing on the LV . 18 All echocardiographic images were saved as DICOM files and analyzed on dedicated 19 workstations. The region of interest was semiautomatically delineated at the end-systolic frame 20 on apical views, which then tracked speckles along the endocardial border throughout the 21 cardiac cycle. The peak negative strain value from each myocardial segment was recorded as 22 the segment’s peak longitudinal systolic strain, and the LVGLS for each image plane was 23 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 10 calculated by averaging the segmental values. The final LVGLS was derived by averaging the 1 LVGLS from the apical 3-chamber (A3C), A4C, and A2C views. For ease of interpretation, the 2 LVGLS values were converted to absolute values. 3 4 2.3. CMR Acquisition and Analysis 5 CMR images were acquired using either a 1.5 -T system (Intera CV release 10; Philips 6 Healthcare, Amsterdam, Netherlands) or a 3 -T system (Philips Ingenia; Best, Netherlands). 7 Imaging was performed under electrocardiographic gating and breath-hold conditions. Steady-8 state free -precession (SSFP) cine -CMR images were obtained in the horizontal long axis, 9 vertical long axis, and LV outflow tract. A short-axis stack view of the whole LV was obtained 10 for LV volume and mass analysis. The cine-CMR sequence was acquired at 25-30 frames/R-R 11 interval (field of view, 320-370 mm; repetition time/echo time, 3.0 -3.6/1.5-1.8ms; flip angle, 12 45-60°; and slice thickness, 6 -8 mm). Late gadolinium enhancement (LGE) images were 13 obtained 10 minutes after intravenous administration of 0.2 mmol/kg of gadodiamide 14 (Omniscan; GE Healthcare, Princeton, NJ, USA) using a phase -sensitive inversion-recovery 15 turbo field echo sequence (repetition/echo time: 4.5 –4.6/1.3–1.5ms, flip angle: 20 –25°, slice 16 thickness: 8 mm). 17 All acquired images were processed using CVI42 software (version 5.10; Circle 18 Cardiovascular Imaging, Calgary, Canada) and analyzed by an independent radiologist blinded 19 to clinical and echocardiographic data. LV volumes (LVEDVCMR and LVESVCMR), LVEFCMR, 20 and LVMCMR were estimated from short-axis cine-CMR. LVGLS analysis was conducted via 21 CMR tissue-tracking (LVGLSCMR-TT), which uses a mid-surface curvilinear coordinate system 22 to track myocardial deformation and follows the motion of software -generated myocardial 23 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 11 nodes on SSFP cine sequences.(18) Endocardial and epicardial borders were traced in a semi-1 automated fashion, with manual correction for contours with apparent deviation. Using long -2 axis cine-images, the whole myocardial LVGLS was calculated throughout the cardiac cycle 3 by the software. As with TTE -derived LVGLS values, the peak negative value (peak systolic 4 strain) was converted to an absolute value, designated as LVGLS CMR-TT. LGE mass was 5 quantified using the full-width at half-maximum method. The total LGE mass was obtained by 6 summing the LGE from all sections, and the relative extent of LGE was expressed as a 7 percentage of total LVM. Extensive LGE was defined as involving >15% of LVM, a threshold 8 associated with an increased risk of sudden cardiac death (SCD).(1, 17, 19) 9 10 2.4. AI-based System for Automatic TTE Analysis 11 Our AI-based system (Sonix Health Workstation, version 2.0; Ontact Health Inc., Korea) 12 perform fully automated tasks, including view classification, segmentation, and 13 echocardiographic parameter extraction. (10, 13 -15) However, since TTE views for GLS 14 measurement were pre-selected for comparison with two conventional LVGLS analysis system 15 in the previous study,(17) view classification was not conducted separately in this study. 16 17 2.5. Segmentation-based Myocardial Advanced Refinement Tracking (SMART) 18 The SMART technique is a novel technique that combines myocardial segmentation with 19 speckle tracking to improve the accuracy of LV quantification in 2D TTE. Integrating structural 20 information allows for precise measurement of important LV functional metrics like LVEDV , 21 LVESV , LVM, and LVGLS. This approach addresses limitations of traditional speckle tracking 22 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 12 methods, which often have difficulty accurately tracking endocardial and epicardial borders in 1 the presence of trabeculations or ultrasound artifacts such as shadows and reverberation. 2 Moreover, traditional 2D speckle tracking often struggles to capture complex 3 -dimensional 3 (3D) motion components, such as longitudinal, circumferential, and radial movements, 4 resulting in errors in strain estimation. 5 Recent advancements in segmentation models that utilize convolutional neural 6 networks (CNNs) and vision transformers (ViTs) allow accurate the LV wall and cavity 7 delineation across diverse patient populations. ( 20-22) These models effectively capture 8 structural nuances and correlate highly with manual LV volume measurements, improving 9 boundary detection and quantification of LVM and function. In the SMART framework, 10 speckle tracking begins with estimated motion vectors, refined using segmentation -based 11 information to better align with myocardial structures. (Figure 1) This refinement involves a 12 two-step process applied to the updated region of interest (ROI) curve, with adjustments made 13 based on a raster scan along its normal direction. In the first step, the ROI's mid -myocardial 14 line is refined to reflect the underlying myocardial structure more accurately. In the second step, 15 the epicardial and endocardial borders are adjusted to ensure precise delineation of the LV wall. 16 This integration enhances the accuracy of endocardial and epicardial border tracking and ROI 17 delineation, resulting in more reliable strain measurement and simultaneous assessment of LV 18 volume and mass. Detailed methodology is provided in Supplemental Method 2. 19 The SMART framework employs advanced tracking and segmentation techniques to 20 accurately calculate LV volume and mass by delineating endocardial and epicardial borders 21 throughout the cardiac cycle. These borders are extracted for every frame in the A4C, A2C, and 22 A3C views. This process generates geometric parameters, such as LV radius and myocardial 23 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 13 thickness, at the basal, mid, and apical levels. LV volumes are computed using a tri -planar 1 configuration (A4C-A2C-A3C), which combines ellipsoid fitting along three axes for precise 2 volume calculations. LV volume is derived from the endocardial border (LVEDV SMART, and 3 LVESVSMART); LVM (LVM SMART) is calculated as the difference between epicardial and 4 endocardial volumes multiplied by a myocardial tissue density of 1.05. This method enables 5 precise LVM estimation through detailed geometrical modelling of the LV wall. 6 In-house validation results of the SMART technique in the internal test datasets, the 7 American population, and a paired dataset with both TTE and CMR are detailed in 8 Supplemental Methods 3-5. 9 10 2.5. Statistical analysis 11 The concordance between LVGLS derived from SMART technique and two different 12 conventional software programs were assessed using the Pearson Correlation Coefficient (PCC) 13 and mean absolute error (MAE). The agreement was further evaluated through Bland-Altman 14 analysis, reporting the mean difference and limits of agreement. Similarly, LVMWT TTE, 15 LVMTTE-DB, and LVMSMART were compared against LVMCMR, the ground truth. Discrimination 16 performance for extensive LGE (>15% LV M) was evaluated across different LVGLS and 17 LVMWT and LVM methods using the area under the receiver operating characteristic curves 18 (AUC). Optimal cutoffs for LVGLS SMART and LVM SMART were obtained from receiver 19 operating characteristics (ROC) curves to predict extensive LGE, enabling patient stratification. 20 Clinical outcomes, including heart failure admission, new -onset atrial fibrillation (AF), and 21 new implantable cardioverter -defibrillators (ICD) insertion, were analyzed as a composite 22 endpoint. Survival curves and Cox regression analysis, adjusted for age, sex, body mass index, 23 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 14 systolic blood pressure, LVEF, and e' velocity, were used to compare outcome risks and 1 calculate hazard ratios (HR). Statistical analyses were performed using R software (version 2 4.3.2; R Development Core Team, Vienna, Austria), with two-sided p-values <0.05 considered 3 statistically significant. 4 5 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 15 3. RESULT 1 3.1. Baseline Characteristics 2 Table 1 summarizes the baseline characteristics of 111 HCM patients (median age: 58 years, 3 male 69%). Among them, 41 (37%) had mixed or diffuse type HCM, 37 (33%) had septal type 4 HCM, and 33 (30%) had apical type HCM. Additionally, 16 patients (14%) exhibited extensive 5 LGE. While age, sex, and BMI did not significantly differ between patients with and without 6 extensive LGE, those with extensive LGE had a higher prevalence of mixed or diffuse HCM 7 phenotypes. 8 LVEDV and LVESV values, regardless of the measurement method, were similar 9 across groups. However, LVEF measured by TTE, whether by conventional manual 10 measurement or the SMART technique, was lower in patients with extensive LGE. Interestingly, 11 LVEF measured by CMR showed no significant difference between the groups. In contrast, 12 LVGLS was significantly reduced in patients with extensive LGE, regardless of the 13 measurement method, indicating more impaired LV function. This reduction was observed with 14 EchoPAC (LVGLSTTE-EchoPAC, 11% vs. 14%, p=0.002) , TomTec (LVGLSTTE-TomTec, 16% vs. 15 20%, p=0.001) , the SMART technique (LVGLSSMART, 11% vs. 14%, p=0.005) , and CMR 16 (LVGLSCMR-TT, 10% vs 11%, p<0.001) . Conventional LVM measures, such as LVMWT TTE 17 and LVMTTE-DB, showed no significant differences between patients with and without extensive 18 LGE. However, LVMSMART was significantly higher in patients with extensive LGE (161 g vs. 19 133 g; p=0.037) as was LVMCMR (188 g vs. 123 g; p=0.007). 20 21 3.2. Performance of SMART Technique 22 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 16 The differences between methods for LVGLS measurement are summarized in Table 2. The 1 correlation plot demonstrates LVGLS SMART is well correlated with LVGLS by conventional 2 vendors. (Figure 2) Specifically, since LVGLS SMART track the mid -myocardium, it showed 3 higher concordance LVGLSTTE-EchoPAC (mean difference, 0 [-1 to 0]; p for difference=0.312). In 4 contrast, LVGLSSMART tended to be lower compared to LVGLS TTE-TomTec, which tracks the 5 endocardium (mean difference, -6 [ -7 to -6]; p for difference <0.001). When compared to 6 LVGLSCMR, all TTE -based LVGLS values showed significant differences (p for difference 7 <0.001); however, LVGLS SMART exhibited the smallest discrepancy (mean difference, 8 LVGLSSMART; 2 [1 to 3], LVGLSTTE_EchoPAC; 3 [2 to 3], and LVGLSTTE_TomTec; 9 [8 to 10]). The 9 correlation plots are provided in Supplemental Result 1. 10 For LVEDV and LVESV , measurements automatically obtained using SMART 11 technique tended to be larger compared to conventional manual measurements in TTE (mean 12 difference, LVEDV; 11 [4 to 19], and LVESV; 7 [4 to 9]). Interestingly, this adjustment reduced 13 the well-known underestimation of LV volumes by TTE compared to CMR (mean difference, 14 LVEDVSMART; -27 [-33 to -21] vs. LVEDVTTE; -37 [-44 to -31], and LVESVSMART; -3 [-6 to -15 0] vs. LVESV TTE; -10 [-13 to -7]), showing slightly improved alignment with CMR values 16 through the SMART technique. As a result, LVEFSMART tended to be slightly lower compared 17 to both conventional manually measured LVEF TTE (mean difference, -3 [ -5 to -2]) and 18 LVEFCMR (mean difference, -7 [-9 to -4]). 19 20 3.3. Feasibility of LVM Measurement by SMART Technique 21 Using LVMCMR as the reference, we examined the correlation of LVMWTTTE and LVMTTE_DB, 22 which are commonly used indices for assessing the severity of LV hypertrophy in conventional 23 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 17 TTE analysis. The correlations are shown in Figure 3. Although LVMTTE-DB showed better 1 correlations with LVMCMR compared to LVMWTTTE, there remained a significant difference 2 (mean difference 106 [90 to 123]; p for difference <0.001). In contrast, unlike LVM TTE-DB, 3 which fails to account for asymmetric hypertrophy, LVMSMART estimates LVM by delineating 4 the epicardial and endocardial borders of the LV wall from the apical views. As a result, it not 5 only demonstrated improved correlation (PCC 0.843 [0.779 to 0.890], p<0.001) but also 6 showed excellent agreement with LVMCMR (mean difference, 1 [-11 to 13], p for difference = 7 0.903). 8 9 3.4. Predictive Value of SMART based LVGLS and LVM for Extensive LGE 10 We performed ROC curve analysis to assess whether extensive LGE (>15% of LVM) observed 11 on CMR could be predictive using LVGLSSMART. The AUC was 0.72 (0.58 to 0.86), comparable 12 to other TTE -based LVGLS values (AUC 0.75 [0.64 to 0.86] for LVGLSTTE-EchoPAC, p for 13 difference =0.708; AUC 0.75 [0.62 to 0.88] for LVGLSTTE-TomTec, p for difference =0.693) and 14 showing no significant difference compared to LVGLS CMR (AUC 0.78 [0.68 to 0.88], p for 15 difference =0.212). (Figure 4) Similarly, we evaluated whether indices used for assessing LV 16 hypertrophy could predict extensive LGE. Conventional TTE -based metrics showed limited 17 predictive ability (AUC 0.59 [0.43 to 0.74] for LVMWTTTE; AUC 0.57 [0.39 to 0.74] for 18 LVMTTE-DB). However, when LVM was measured using the SMART technique, the AUC 19 increased to 0.66 (0.51 to 0.82), which was comparable to LVM CMR with no significant 20 difference observed (AUC 0.71 [0.56 to 0.86], p for difference =0.130). 21 22 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 18 3.5. Outcome Risk Stratification Based on Extensive LGE Predicted by Auto -Measured 1 LVGLS and LVM 2 The optimal cutoff values for extensive LGE were identified as 11% for LVGLSSMART and 166 3 g for LVMSMART. Patients categorized as high -risk for extensive LGE based on LVGLS SMART 4 (≤11%) had a significantly higher risk of adverse clinical outcomes, with an adjusted HR of 5 6.02 (2.21 to 16.41). ( Figure 5) Similarly, those classified as high -risk based on LVM SMART 6 (≥166 g) also demonstrated an increased risk of clinical outcomes, with an adjusted HR of 3.97 7 (1.50 to 10.47). 8 9 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 19 4. DISCUSSION 1 In this study, we introduce the SMART technique, a novel approach that refines motion tracking 2 by incorporating LV myocardial segmentation throughout the cardiac cycle. We demonstrated 3 that the SMART technique is feasible and effectively segment s LV walls in HCM patients, 4 enabling accurate assessment of LVM and LVGLS. For predicting extensive LGE, 5 LVGLSSMART showed predictive value comparable to LVGLS measured by well-stablished 6 systems, while LVMSMART demonstrated superior predictive value compar ed to conventional 7 TTE measures such as LVMWTTTE or LVMTTE-DB. Additionally, we confirmed the independent 8 prognostic value of both LVGLSSMART and LVMSMART. 9 Conventional speckle tracking system often relies on semi-automated, operator -10 defined endocardial border delineation, whether tracking the endocardium or mid-myocardium. 11 While effective for assessing global LV motion, this approach does not precisely define both 12 endocardial and epicardial borders to establish an ROI for tracking. Recently, AI-based 13 automatic LV strain analysis has been introduced, enabling automatic segmentation and 14 tracking of the LV wall. (8, 14, 22) Salte et al. demonstrated the feasibility and accuracy of 15 automated LVGLS measurement in 200 patients with various cardiovascular pathologies and 16 levels of LV function, using DL-based LV wall segmentation and mid-wall tracking.(8) They 17 later reported improved precision and reproducibility with this approach .(22) Another study 18 involving 870 COVID -19 patients showed that AI-derived LVEF and LVGLS, based on LV 19 endocardial border tracing, had superior accuracy in predicting mortality compared to manual 20 measurement.(23) We recently presented a system that automatic segment s the LV wall to 21 define and track the mid-wall, enabling accurate LVEF and LVGLS measurement.(14) In 22 revascularized STEMI patients, our system accurately measured LVEF and LVGLS compared 23 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 20 to manual measurements and demonstrated comparable prognostic value. However, most of 1 these studies focused on patients with varying levels of LV dysfunction and did not specifically 2 address LV hypertrophy. The ability of DL -based techniques to accurately delineate the 3 endocardial and epicardial borders and track LV myocardial motion in HCM patients, 4 characterized by severe and asymmetric LV hypertrophy, remains insufficiently validated. 5 Accurate tracking of LV wall motion requires precise and automated segmentation, 6 which is particularly challenging in TTE due to lower tissue contrast, susceptibility to artifacts, 7 and limited resolution compared to CMR. This aligns with the well -documented challenge of 8 accurately quantifying LV wall hypertrophy in HCM using TTE . For instance, prior studies 9 have reported discrepancies between LVMWT measured by CMR and TTE.(24, 25) Moreover, 10 conventional TTE-derived LVM, which relies on geometric assumptions, often fails to capture 11 asymmetric hypertrophy pattern, such as those seen in septal or apical HCM, as demonstrated 12 in the present study. In contrast, t he SMART technique addresses these limitations by 13 integrating accurate LV wall segmentation from A4C, A2C, and A3C views to refine speckle 14 tracking. This approach provides a more precise representation of LV structure, even in cases 15 with asymmetric geometry, resulting in LVMSMART showing high concordance with LVMCMR. 16 Given the critical prognostic value of LVM in HCM, the ability to automatically and accurately 17 measure LVM using TTE carries significant clinical implications. In our study, LVM SMART, 18 unlike LVMWTTTE or LVMTTE-DB, successfully predicted extensive LGE and demonstrated 19 independent prognostic value. 20 The SMART technique's precise LV segmentation and integration of tri-plane imaging 21 improved the accuracy of LV volume measurements, reduc ing discrepancies with CMR -22 derived volumes compared to conventional bi -plane assessment . While Simpson’s biplane 23 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 21

is more accurate than dimension -based approaches, it still relies on geometric 1 assumptions about LV shape (2), which are well-documented to result in 2D TTE 2 underestimating LV volumes compared to CMR. (26, 2 7) These assumptions introduce 3 inaccuracies in cases with anatomical variability , such as slanted basal segments, cross -4 sectional ellipticity, or foreshorten ed planes.(28) These challenges are further amplified i n 5 HCM due to asymmetric hypertrophy. Although TTE-based measurement of LV volume, 6 including the SMART technique , has inherent limitations compared to CMR, the SMART 7 technique’s enhanced LV wall segmentation and tri -plane approach offer significant 8 improvements over conventional 2D TTE analysis. 9 Even with fully automated LV segmentation without manual intervention, the SMART 10 technique demonstrated high accuracy for measuring LVGLS. By performing segmentation 11 across all frames, rather than being limited to ES, SMART refines motion tracking and enables 12 precise LVGLS measurement. The SMART-derived LVGLS showed strong concordance with 13 LVGLS manually measured by experts using well -established systems. Notably, since the 14 SMART technique employs mid-wall tracking, its results align more closely with those from 15 the EchoPAC system, which also uses mid -wall tracking, than with TomTec system, which 16 relies on endocardial tracking. The concordance with CMR was lower, as expected, due to 17 differences in imaging modalities and examination dates. While TTE-derived LVGLS 18 measurements were obtained from the same views of the same echocardiographic exam using 19 different methods, LVGLSCMR-TT was measured on the cine-CMR conducted on a separate day 20 with non -identical imaging planes. Despite these differences, LVGLS SMART demonstrated 21 predictive ability for extensive LGE comparable to LVGLS CMR and also showed independent 22 prognostic value. 23 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 22 Despite these promising results, several limitations need to be acknowledged. 1 Although this study demonstrated the feasibility of the SMART technique in HCM, the sample 2 size was relatively small. Nevertheless, the study cohort included a unique structure 3 incorporating LV geometric and three different LVGLS parameters derived from both TTE and 4 CMR, allowing for a robust comparison of SMART's performance against these imaging 5 modalities. Future research involving a more diverse population of cardiomyopathy subtypes, 6 beyond HCM, is necessary to further validate the diagnostic feasibility and clinical 7 applicability of the SMART technique. 8 9 5. CONCLUSION 10 The SMART technique is an AI-based novel technique that combines LV speckle tracking with 11 myocardial segmentation in a hybrid approach based on TTE. In patients with HCM, 12 characterized by LV hypertrophy with asymmetric geometry, the SMART technique allows for 13 comparable evaluation of LVGLS and more accurate assessment of LV M compared to 14 conventional measurements in TTE. Furthermore, SMART -derived LVGLS and LV M 15 demonstrated utility in predicting myocardial fibrosis and clinical outcomes, supporting the 16 diagnostic and prognostic feasibility of this novel technique in the management of HCM. 17 18 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 23 Contributors 1 All authors contributed equally to this study. All authors have read and approved the final 2 version of the manuscript. J.P, J.J, and Y .E.Y verified the underlying data of the current study. 3 4 Data Sharing Statement 5 Data cannot be made publicly available due to ethical restrictions set by the IRB of the study 6 institution; i.e., public availability would compromise patient confidentiality and participant 7 privacy. Please contact the corresponding author ([email protected]) to request the 8 minimal anonymi sed dataset. Researchers with additional inquiries about the deep learning 9 model presented in this study are also encouraged to reach out to the corresponding author. 10 11 Declaration of Interests 12 Y .E.Y , Y .J., Y .H., and S.A.L. are currently affiliated with Ontact Health, Inc. H.J.C. holds stock 13 in Ontact Health, Inc. The other authors have no conflicts of interest to declare. 14 15 Funding 16 This work was supported by a grant from the Institute of Information & communications 17 Technology Planning & Evaluation (IITP) funded by the Korea government (Ministry of 18 Science and ICT) (No.2022000972, Development of a Flexible Mobile Healthcare Software 19 Platform Using 5G MEC); and the Medical AI Clinic Program through the NIPA funded by the 20 MSIT. (Grant No.: H0904-24-1002). 21 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 24

1 1. Ommen SR, Ho CY , Asif IM, Balaji S, Burke MA, Day SM, et al. 2024 2 AHA/ACC/AMSSM/HRS/PACES/SCMR Guideline for the Management of Hypertrophic 3 Cardiomyopathy: A Report of the American Heart Association/American College of 4 Cardiology Joint Committee on Clinical Practice Guidelines. Circulation. 2024;149(23):e1239-5 e311. 6 2. Lang RM, Badano LP, Mor -Avi V , Afilalo J, Armstrong A, Ernande L, et al. 7 Recommendations for cardiac chamber quantification by echocardiography in adults: an update 8 from the American Society of Echocardiography and the European Association of 9 Cardiovascular Imaging. J Am Soc Echocardiogr. 2015;28(1):1-39 e14. 10 3. Devereux RB, Alonso DR, Lutas EM, Gottlieb GJ, Campo E, Sachs I, et al. 11 Echocardiographic assessment of left ventricular hypertrophy: comparison to necropsy 12 findings. Am J Cardiol. 1986;57(6):450-8. 13 4. Kristensen CB, Myhr KA, Grund FF, Vejlstrup N, Hassager C, Mattu R, et al. A new 14

to quantify left ventricular mass by 2D echocardiography. Sci Rep. 2022;12(1):9980. 15 5. Zhang J, Gajjala S, Agrawal P, Tison GH, Hallock LA, Beussink-Nelson L, et al. Fully 16 Automated Echocardiogram Interpretation in Clinical Practice. Circulation. 17 2018;138(16):1623-35. 18 6. Yoon YE, Kim S, Chang HJ. Artificial Intelligence and Echocardiography. J 19 Cardiovasc Imaging. 2021;29(3):193-204. 20 7. Zhou J, Du M, Chang S, Chen Z. Artificial intelligence in echocardiography: detection, 21 functional evaluation, and disease diagnosis. Cardiovasc Ultrasound. 2021;19(1):29. 22 8. Salte IM, Ostvik A, Smistad E, Melichova D, Nguyen TM, Karlsen S, et al. Artificial 23 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 25 Intelligence for Automatic Measurement of Left Ventricular Strain in Echocardiography. JACC 1 Cardiovasc Imaging. 2021;14(10):1918-28. 2 9. Kwan AC, Chang EW, Jain I, Theurer J, Tang X, Francisco N, et al. Deep Learning -3 Derived Myocardial Strain. JACC Cardiovasc Imaging. 2024;17(7):715-25. 4 10. Jeon J, Ha S, Yoon YE, Kim J, Jeong H, Jeong D, et al. Echocardiographic view 5 classification with integrated out -of-distribution detection for enhanced automatic 6 echocardiographic analysis. arXiv [eess.SP; 2023]. Available from: 7 https://arxiv.org/abs/2308.16483. 8 11. Jeon J, Kim J, Jang Y , Yoon YE, Jeon D, Hong Y , et al. A Unified Approach for 9 Comprehensive Analysis of Various Spectral and Tissue Doppler Echocardiography. arXiv 10 [eess.IV; 2023]. Available from: https://arxiv.org/abs/2311.08439. 11 12. Jeong D, Jung S, Yoon YE, Jeon J, Jang Y , Ha S, et al. Artificial intelligence-enhanced 12 automation for M -mode echocardiographic analysis: ensuring fully automated, reliable, and 13 reproducible measurements. Int J Cardiovasc Imaging. 2024;40(6):1245-56. 14 13. Park J, Jeon J, Yoon YE, Jang Y , Kim J, Jeong D, et al. Artificial intelligence-enhanced 15 automation of left ventricular diastolic assessment: a pilot study for feasibility, diagnostic 16 validation, and outcome prediction. Cardiovascular Diagnosis and Therapy. 202 4;14(3):352-17 66. 18 14. Jang Y , Choi H, Yoon YE, Jeon J, Kim H, Kim J, et al. An Artificial Intelligence-Based 19 Automated Echocardiographic Analysis: Enhancing Efficiency and Prognostic Evaluation in 20 Patients With Revascularized STEMI. Korean Circ J. 2024;54(11):743-56. 21 15. Park J, Kim J, Jeon J, Yoon YE, Jang Y , Jeong H, et al. Artificial Intelligence-Enhanced 22 Comprehensive Assessment of the Aortic Valve Stenosis Continuum in Echocardiography. 23 medRxiv 2024.07.08.24310123. Available from: https://doi.org/10.1101/2024.07.08.24310123. 24 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 26 16. Agency NIS. Open AI Dataset Project (AI-Hub) [Available from: https://aihub.or.kr/. 1 17. Park J, Yoon YE, Chun EJ, Choi HM, Hwang IC, Lee HJ, et al. Endocardial versus 2 whole-myocardial tracking global longitudinal strain analysis in patients with hypertrophic 3 cardiomyopathy: A preliminary comparative study. PLoS One. 2023;18(7):e0288421. 4 18. Yoon YE, Kang SH, Choi HM, Jeong S, Sung JM, Lee SE, et al. Prediction of infarct 5 size and adverse cardiac outcomes by tissue tracking -cardiac magnetic resonance imaging in 6 ST-segment elevation myocardial infarction. Eur Radiol. 2018;28(8):3454-63. 7 19. Chan RH, Maron BJ, Olivotto I, Pencina MJ, Assenza GE, Haas T, et al. Prognostic 8 value of quantitative contrast-enhanced cardiovascular magnetic resonance for the evaluation 9 of sudden death risk in patients with hypertrophic cardiomyopathy. Circulation. 10 2014;130(6):484-95. 11 20. Duffy G, Cheng PP, Yuan N, He B, Kwan AC, Shun-Shin MJ, et al. High-Throughput 12 Precision Phenotyping of Left Ventricular Hypertrophy With Cardiovascular Deep Learning. 13 JAMA Cardiol 2022;7(4):386–95. 14 21. Wei H, Ma J, Zhou Y , Xue W, Ni D. Co-learning of appearance and shape for precise 15 ejection fraction estimation from echocardiographic sequences. Med Image Anal 16 2023;84:102686. 17 22. Christensen M, Vukadinovic M, Yuan N, Ouyang D. Vision -language foundation 18 model for echocardiogram interpretation. Nat Med. 2024;30(5):1481-8. 19 22. Salte IM, Ostvik A, Olaisen SH, Karlsen S, Dahlslett T, Smistad E, et al. Deep Learning 20 for Improved Precision and Reproducibility of Left Ventricular Strain in Echocardiography: A 21 Test-Retest Study. J Am Soc Echocardiogr. 2023;36(7):788-99. 22 23. Asch FM, Descamps T, Sarwar R, Karagodin I, Singulane CC, Xie M, et al. Human 23 versus Artificial Intelligence-Based Echocardiographic Analysis as a Predictor of Outcomes: 24 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 27 An Analysis from the World Alliance Societies of Echocardiography COVID Study. J Am Soc 1 Echocardiogr. 2022;35(12):1226-37.e7. 2 24. Urbano-Moral JA, Gonzalez-Gonzalez AM, Maldonado G, Gutierrez-Garcia-Moreno 3 L, Vivancos-Delgado R, De Mora -Martin M, et al. Contrast-Enhanced Echocardiographic 4 Measurement of Left Ventricular Wall Thickness in Hypertrophic Cardiomyopathy: 5 Comparison with Standard Echocardiography and Cardiac Magnetic Resonance. J Am Soc 6 Echocardiogr. 2020;33(9):1106-15. 7 25. Hindieh W, Weissler-Snir A, Hammer H, Adler A, Rakowski H, Chan RH. Discrepant 8 Measurements of Maximal Left Ventricular Wall Thickness Between Cardiac Magnetic 9 Resonance Imaging and Echocardiography in Patients With Hypertrophic Cardiomyopathy. 10 Circ Cardiovasc Imaging. 2017;10(8):e006309. 11 26. Rigolli M, Anandabaskaran S, Christiansen JP, Whalley GA. Bias associated with left 12 ventricular quantification by multimodality imaging: a systematic review and meta -analysis. 13 Open Heart. 2016;3(1):e000388. 14 27. Manole S, Budurea C, Pop S, Iliescu AM, Ciortea CA, Iancu SD, et al. Correlation 15 between V olumes Determined by Echocardiography and Cardiac MRI in Controls and Atrial 16 Fibrillation Patients. Life (Basel). 2021;11(12):1362. 17 28. Kim WC, Beqiri A, Lewandowski AJ, Puyol-Anton E, Markham DC, King AP, et al. 18 Beyond Simpson's Rule: Accounting for Orientation and Ellipticity Assumptions. Ultrasound 19 Med Biol. 2022;48(12):2476-85. 20 21 22 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 28 Figure Legends 1 Figure 1. Workflow of the SMART Framework: Integrating Motion Tracking and 2 Segmentation for LV Quantification 3 The workflow highlights the dynamic update of LV ROI using a bi-directional dense motion 4 field and a segmentation-guided refinement process to achieve myocardial structure -aware 5 tracking for enhanced functional assessment. The process begins with segmentation of the LV 6 myocardium across all frames, constructing an LV area curve for cardiac cycle analysis, 7 including the identification of key phases such as ED and ES. Motion vectors are estimated 8 and iteratively refined through a two -step process: first, by aligning the mid -myocardial line 9 with the myocardial structure, and second, by precisely delineating the epicardial and 10 endocardial borders. 11 Abbreviations: ED, end-diastole; ES, end -systole; LV , left ventricle; ROI, region of interest; 12 SMART, Segmentation-based Myocardial Advanced Refinement Tracking. 13 14 Figure 2. Concordance of LVGLS between SMART and manual measurements 15 The correlation plot demonstrated a significant association between SMART and manual 16 measurements for LVGLS using conventional vendors (EchoPAC (A) and TomTec (B)). 17 Abbreviations as in Figure 1: LVGLS, left ventricular global longitudinal strain; MAE, mean 18 absolute error; PCC, Pearson Correlation Coefficient; TTE, transthoracic echocardiography. 19 20 Figure 3. Concordance of SMART and conventional dimension -based measurement to 21 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 29 CMR for LVM 1 Compared to LVMWT (A) and dimension-based LVM (B) measured by TTE, SMART-based 2 LVM (C) showed stronger concordance LVM assessed by CMR. 3 Abbreviations as in Figure 1 and 2: CMR, cardiovascular magnetic resonance imaging; DB, 4 dimension-based methods; LVM, left ventricular mass; LVMWT, left ventricular maximal wall 5 thickness. 6 7 Figure 4. Predictive value of SMART-based LVM and LVGLS for extensive LGE 8 Compared to conventional vendors, SMART-based LVGLS showed comparable performance 9 for predicting extensive LGE (A). SMART-based LVM outperformed LVMWT and dimension-10 based LVM for predicting extensive LGE (B). 11 Abbreviations as in Figure 1-3: AUC, area under the receiver operating characteristic curves; 12 CI, confidence interval 13 14 Figure 5. Outcome Risk Stratification Based on SMART-based LVM and LVGLS 15 Stratification based on SMART-based LVGLS (≤11%) (A) and LVM (≥166g) (B) identified 16 high-risk patients for composite outcomes. 17 Abbreviations as in Figure 1-4: HR, hazard ratio 18 19 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint 30 Supplemental Methods 1 Supplemental Method 1. Dataset for the Development of the SMART Technique 2 Supplemental Method 2. Technical Detail of the SMART Technique 3 Supplemental Method 3. In-House Validation of SMART-Derived LV Volume and GLS 4 Measurements 5 Supplemental Method 4. External Validation of the SMART-Derived GLS Measurement 6 in American Population 7 Supplemental Method 5. Comparison of SMART Technique with Conventional TTE 8 and CMR Measurement in a Dataset with Paired Echocardiography and CMR Imaging 9 10 Supplemental Results 11 Supplemental Result 1. Concordance of SMART and conventional vendors to CMR for 12 LVGLS 13 Abbreviations: CMR, cardiovascular magnetic resonance imaging; LVGLS, left ventricular 14 global longitudinal strain; MAE, mean absolute error; PCC, Pearson Correlation Coefficient; 15 SMART, Segmentation-based Myocardial Advanced Refinement Tracking; TTE, 16 transthoracic echocardiography. 17 18 All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint All rights reserved. No reuse allowed without permission. perpetuity. preprint (which was not certified by peer review) is the author/funder, who has granted medRxiv a license to display the preprint in The copyright holder for thisthis version posted January 17, 2025. ; https://doi.org/10.1101/2025.01.17.25320694doi: medRxiv preprint