Left ventricular segmental strain and the prediction of cancer therapy-related cardiac dysfunction

Abstract

Aims

We aimed to determine the early changes and predictive value of left ventricular (LV) segmental strain measures in women with breast cancer receiving doxorubicin.

Methods and results

In a cohort of 237 women with breast cancer receiving doxorubicin with or without trastuzumab, 1151 echocardiograms were prospectively acquired over a median (Q1–Q3) of 7 (2–24) months. LV ejection fraction (LVEF) and 36 segmental strain measures were core lab quantified. A supervised machine learning (ML) model was then developed using random forest regression to identify segmental strain measures predictive of nadir LVEF post-doxorubicin completion. Cancer therapy-related cardiac dysfunction (CTRCD) was defined as a ≥10% absolute LVEF decline pre-treatment to a value <50%. Median (Q1–Q3) baseline age was 48 (41–57) years. Thirty-five women developed CTRCD, and eight of these developed symptomatic heart failure. From pre-treatment to doxorubicin completion, longitudinal strain worsened across the basal and mid-LV segments but not in the apical segments; circumferential strain worsened primarily in the septum; radial strain worsened uniformly and transverse strain remained unchanged across all LV segments. In the ML model, anterolateral and inferoseptal circumferential strain were the most predictive features; longitudinal and transverse strain in the basal inferoseptal, anterior, basal anterolateral, and apical lateral segments were also top predictive features. The addition of predictive segmental strain measures to a model including age, cancer therapy regimen, hypertension, and LVEF increased the area under the curve (AUC) from 0.70 (95% confidence interval (CI) 0.60–0.80) to 0.87 (95% CI 0.81–0.92), ΔAUC = 0.18 (95% CI 0.08–0.27) for the prediction of CTRCD.

Conclusion

Our findings suggest that segmental strain measures can enhance cardiotoxicity risk prediction in women with breast cancer receiving doxorubicin.

Introduction

Doxorubicin is a highly effective cancer therapy agent that is commonly used in the treatment of breast cancer. However, the risk of cardiotoxicity leading to cardiac dysfunction and symptomatic heart failure (HF) remains an important concern.^1,2 According to the American Heart Association/American College of Cardiology guideline, patients receiving doxorubicin are considered as having stage A HF, a designation highlighting an increased risk of progression to more advanced stages of HF.³

The development of risk prediction tools for the early identification of patients at greater risk of cancer therapy-related cardiac dysfunction (CTRCD), and HF from cardiotoxic cancer therapy is an area of active research in cardio-oncology.⁴ These tools can enhance personalized decision-making regarding risks and benefits, and can facilitate the use of targeted cardioprotective strategies to prevent progression to advanced stages of HF.^5–7 Strain imaging by speckle tracking echocardiography shows promise as a sensitive tool for the detection of subclinical left ventricular (LV) dysfunction and the prediction of subsequent CTRCD. Strain measurements are obtained across multiple LV segments and dimensions including longitudinal, circumferential, transverse, and radial.⁸ Currently, average measures such as global longitudinal and circumferential strain have primarily been studied as predictors of CTRCD.^7,9–12 The potential value of the data derived from segmental measures of strain remains unknown.

Machine learning (ML) tools are well suited to assess the predictive value of segmental strain measures. These tools are being increasingly investigated for automated image segmentation and quantitation, interpretation, and diagnostics.^13,14 In addition, ML tools have been successfully utilized to identify features that are predictive of outcomes with the development of high performing risk prediction models using data from different cardiovascular imaging modalities.^15,16 However, the application of ML in the field of cardio-oncology has been limited to date.

Here, we applied a supervised ML approach to a dataset of 237 women with quantified two-dimensional LV segmental strain measures to inform the prediction of CTRCD in breast cancer. We specifically aimed to: (i) determine the early changes in segmental strain measures, (ii) identify segmental strain measures that are predictive of LV ejection fraction (LVEF) utilizing supervised ML, and (iii) determine the incremental value of segmental strain measures for the prediction of CTRCD in women with breast cancer receiving doxorubicin-based cancer therapy regimen.

Methods

Study population

The study population is a subcohort of the Cardiotoxicity of Cancer Therapy (CCT) study, a prospective cohort of women with breast cancer from the Rena Rowan Breast Center of the Abramson Cancer Center at the University of Pennsylvania (Philadelphia, PA, USA) that has been previously described.^11,12 Women 18 years of age and older with breast cancer who were treated with doxorubicin with or without trastuzumab were included. Exclusion criteria included pregnancy or an inability or unwillingness to provide informed consent. Treatment regimen was determined by the treating oncologist and consisted of either doxorubicin (240 mg/m² divided into four cycles of 60 mg/m² each) and cyclophosphamide followed by a taxane-containing regimen (Doxorubicin) or doxorubicin (240 mg/m² divided into four cycles of 60 mg/m² per cycle) and cyclophosphamide followed by a taxane and trastuzumab (Doxorubicin + Trastuzumab). The study was performed in accordance with the Declaration of Helsinki and was approved by the Institutional Review Board of the University of Pennsylvania. All women provided written informed consent. The current analysis included 237 women with available study visits pretreatment and at the completion of doxorubicin. The data underlying this article cannot be shared publicly.

Study procedures

The study procedures have been previously described.¹² In brief, clinical data were collected using standardized questionnaires before treatment initiation (i.e. baseline) and during prespecified follow-up visits. Dedicated sonographers at an Intersocietal Accreditation Commission laboratory performed transthoracic echocardiograms. Echocardiography was performed at baseline, at the completion of a taxane (∼4 months), and then annually in the Doxorubicin group. In the Doxorubicin + Trastuzumab group, echocardiography was performed at baseline, at the completion of doxorubicin (∼2 months), every 3 months during trastuzumab therapy, and then annually. Two-dimensional images were acquired using Vivid 7, E9, or E95 machines (GE Healthcare, Milwaukee, WI, USA).

Quantitative echocardiography measures

Quantitative echocardiography analysis was performed from digitally archived images at the University of Pennsylvania Center for Quantitative Echocardiography (Philadelphia, PA, USA). Quantitation of LV volumes and strain were performed using the TomTec Imaging Systems platform (Unterschleissheim, Germany). Simpson’s method of discs was used to derive LVEF. Strain measurements were performed by a single sonographer from images obtained in the apical four-chamber and two-chamber views, and parasternal short-axis view (SAX) at the mid-papillary level. Segmental strain measurements were derived through the same process typically used to obtain global strain measurements. Image quality assessments for apical four- and two-chamber views as well as SAX are provided for the subset of the analysed images in Supplementary data online, Table S1. The LV endocardial border was manually traced at the end-systolic frame of one cardiac cycle. Peak systolic longitudinal and transverse strain were automatically quantified across the apical (inferior, septal, anterior, and lateral), mid (inferior, inferoseptal, anterior, and anterolateral) and basal (inferior, inferoseptal, anterior, and anterolateral) LV segments of the apical four-chamber and two-chamber views. Peak circumferential and radial strain were quantified across the mid (inferior, inferoseptal, inferolateral, anterior, anteroseptal, and anterolateral) LV segments of the parasternal SAX view. Estimates of intraobserver coefficients of variation for the reproducibility of individual segmental strain measures are provided in Supplementary data online, Table S2. In this study, we focused on segmental strain measurements from echocardiograms acquired at baseline and at the completion of doxorubicin. The number of available measurements of each segmental strain measure at the two timepoints is provided in Supplementary data online, Table S3. There were greater percentages of missing segmental strain data (i.e. ∼5–15%) at baseline in comparison with the doxorubicin completion timepoint due to limitations in image quality leading to poor tracking. This was partly attributable to differences in clinical status, including echocardiograms being performed shortly after left-sided mastectomy in a subset of the women (Supplementary data online, Table S4A–C). Quantitated LVEF values from echocardiograms acquired at post-doxorubicin completion follow-up echocardiography visits were used to define outcomes.

Outcomes

We evaluated two outcomes including (i) nadir LVEF after the completion of doxorubicin, defined as the lowest LVEF value during subsequent follow-up after the doxorubicin completion timepoint, and (ii) CTRCD, defined as a ≥10% absolute decline in LVEF from baseline to a value <50% with or without HF symptoms.¹²

Data analysis

Baseline characteristics were summarized using proportions for categorical variables, mean (standard deviation (SD)) for normally distributed continuous variables and median (Q1–Q3) for non-normally distributed continuous variables. The mean (95% confidence interval (CI)) of the difference between the baseline and doxorubicin completion values was calculated for each segmental strain measure, and standardized estimates are presented using the SD of the baseline values. This facilitated the comparison of the magnitude of change across all measures. Differences in the distributions of the segmental strain measures at the two timepoints were determined using the paired t-test.

A supervised ML model was trained to identify LV segmental strain measures at doxorubicin completion that were predictive of nadir LVEF. Given prior work that suggested the prognostic significance of cardiac strain at doxorubicin completion, we utilized segmental strain measures at this timepoint for model training.¹⁰ Missing values were imputed by carrying forward values at the earliest of any prior visits when available. Random forest regression was implemented for model training. Features were standardized during model training. Model parameter tuning was performed using 10 repeats of 10-fold cross-validation. In 10-fold cross-validation, the dataset was randomly split into 10 approximately equal subgroups. Model training was performed in 9/10 of the subgroups, and performance was tested on the remaining subgroup. This process was continued until all the subgroups were used for model testing; each data point was used 9 times for training and once for performance testing (Figure 1). The outputs across the 10 repeats were combined to provide a measure of overall model performance. The model with the combination of parameters that yielded the greatest overall performance was identified as the final model. To identify predictive features and determine relative feature importance, a cross-validated variable importance metric, change in the mean square error (MSE), was calculated using a permutation-based variable importance assessment (Supplementary data online, Methods).¹⁷ The MSE indicates the average of the squared difference between predicted and observed values; therefore, the more positive the change in MSE, the greater the feature importance. Features that resulted in a change in MSE >0 were considered predictive.

Figure 1.

Overview of model training using 10-fold cross-validation. The figure represents the overview of model training with 10-fold cross-validation using 36 left ventricular segmental strain measures for the prediction of nadir LVEF post-doxorubicin completion; the predictive features are presented in the order of the magnitude of feature importance in the machine learning model.

Next, we performed receiver operating characteristic curve analysis to evaluate the incremental value of LV segmental strain measures for the prediction of CTRCD, when added to a reference model including age, cancer therapy regimen, hypertension, and LVEF at doxorubicin completion. Here, we included the segmental strain measures that were determined to be predictive in the ML model for the nadir LVEF outcome. A separate random forest model was not trained for the CTRCD outcome due to the limited number of events. Incremental value was also compared to the gains attained with average longitudinal, circumferential, transverse, and radial strain measures. Bootstrap resampling (N = 1000) was incorporated to calculate and compare performance metrics. We also performed a sensitivity analysis using an alternative definition of CTRCD based on the criteria proposed by the American Society of Echocardiography (ASE) and European Association of Cardiovascular Imaging (EACVI) consensus group, which defines CTRCD as a ≥10% absolute decline in LVEF from baseline to a value <53%.¹⁸ All analyses were performed using R 3.4.0 (R Foundation for Statistical Computing, Vienna, Austria).

Results

Study population

Baseline characteristics of the study population are presented in Table 1. Median (Q1–Q3) age was 48 (41–57) years and 66.2% of included women were Caucasian. The majority (80.6%) were treated with doxorubicin, while 19.4% received doxorubicin followed by trastuzumab. The median (Q1–Q3) LVEF was 54% (51–57) at baseline and 52% (48–55) at doxorubicin completion. Average measures of longitudinal, circumferential, transverse, and radial strain are summarized in Supplementary data online, Table S5.

Table 1.

Baseline characteristics of the study population

Variables	Overall population (N = 237)
Age (years)	48 (41–57)
Race
Black	65 (27.4)
Caucasian	157 (66.2)
Other	15 (6.3)
Breast cancer side
Left	109 (46.2)
Right	112 (47.4)
Bilateral	15 (6.4)
Breast cancer stage
Stage 1	35 (14.8)
Stage 2	142 (59.9)
Stage 3	59 (24.9)
Stage 4	1 (0.4)
Radiation therapy
None	82 (34.7)
Left-sided	76 (32.2)
Right-sided	69 (29.2)
Bilateral	9 (3.8)
Cancer therapy regimen
Doxorubicin	191 (80.6)
Doxorubicin + trastuzumab	46 (19.4)
Left ventricular ejection fraction (%)	54 (51–57)
Body mass index (kg/m2)	26 (23–31)
Systolic blood pressure (mmHg)	124 (112–132)
Diastolic blood pressure (mmHg)	74 (69–81)
Heart rate (bpm)	80 (73–89)
Current or past smoking	90 (38.6)
History of diabetes mellitus	18 (7.6)
History of hypertension	64 (27.1)
History of hyperlipidaemia or statin use	53 (22.4)
Hyperlipidaemia	52 (21.9)
Statin use	23 (9.7)
ACEI	19 (8.0)
ARB	13 (5.5)
Beta-blocker	19 (8.0)

Count (%) is presented for categorical variables; median (Q1–Q3) is presented for continuous variables.

ACEI, angiotensin converting enzyme inhibitor; ARB, angiotensin receptor blocker; none of the women had history of myocardial infarction or coronary heart disease at baseline.

Early changes in LV segmental strain measures

Overall, the worsening in longitudinal strain from baseline to the completion of doxorubicin was more pronounced in the basal and mid-LV segments, when compared with the apical segments (Figure 2A, Supplementary data online, Table S6). The worsening in longitudinal strain was greater in the basal inferoseptal, basal anterior, and mid anterolateral segments. In the basal inferoseptal segment, longitudinal strain worsened from a mean (SD) of −13.8% (5.5) at baseline to −12.4% (5.7) at doxorubicin completion (P = 0.001), while longitudinal strain worsened from −21.0% (6.5) to −19.1% (6.5) (P = 0.016) in the basal anterior, and −15.6% (5.1) to −14.1% (4.8) (P = 0.002) in the mid anterolateral segments. Interestingly, longitudinal strain was largely preserved across the apical LV segments. Transverse strain did not significantly change from baseline to doxorubicin completion across all apical, mid and basal LV segments (Figure 2B, Supplementary data online, Table S7). However, there was a suggestion of an increase in transverse strain in the apical lateral and mid anterolateral segments with average increases from 23.9% (SD = 18.7) to 27.0% (SD = 21.0) (P = 0.062) and 45.9% (SD = 35.5) to 50.0% (SD = 36.7) (P = 0.152), respectively, although these were not statistically significant.

Figure 2.

Early changes in LV segmental strain measures from pre-treatment to doxorubicin completion. The figure summarizes standardized mean change from pre-treatment to doxorubicin completion in longitudinal (A), transverse (B), circumferential (C), and radial (D) LV segmental strain. A more positive change in longitudinal and circumferential strain indicates worsening function; a more negative change in transverse and radial strain indicates worsening function.

Circumferential strain did not significantly change from baseline with the exception of the septal segments, particularly the mid inferoseptal segment (Figure 2C, Supplementary data online, Table S8). Mid inferoseptal circumferential strain worsened from a mean (SD) of −28.9% (7.3) at baseline to −27.4% (8.3) at doxorubicin completion (P = 0.004). There was also a significant worsening of radial strain in this segment. The mean (SD) mid inferoseptal radial strain decreased from 39.7% (24.4) to 35.3% (24.7) (P = 0.021). A significant reduction in radial strain was similarly observed in the mid inferior segment; radial strain decreased from a baseline mean (SD) value of 60.2% (36.0) to 51.2% (30.7) (P = 0.004) doxorubicin completion (Figure 2D, Supplementary data online, Table S9).

Comparative analysis of the predictive value of LV segmental strain measures

This analysis was performed in 202 women with available LVEF at doxorubicin completion and during at least one subsequent follow-up visit. Baseline characteristics of this subcohort are presented in Supplementary data online, Table S10. A total of 845 quantified LVEF values were available in this subset over a median (Q1–Q3) follow-up period of 10 (2–26) months. The median (Q1–Q3) nadir LVEF during the post-doxorubicin completion follow-up period was 49% (44–52). CTRCD occurred in 35 women during this follow-up period, and out of these eight developed New York Heart Association class II HF. Baseline characteristics of these women are summarized in Supplementary data online, Table S11. Overall, women who developed CTRCD were more frequently treated with sequential doxorubicin and trastuzumab therapy and had higher prevalence of cardiovascular risk factors at baseline.

Among the 36 LV segmental strain measures available for model training at the doxorubicin completion timepoint, mid anterolateral and inferoseptal circumferential strain were identified as the most predictive features for nadir (Figure 3, Supplementary data online, Table S12). Moreover, longitudinal strain measures across the inferoseptal (basal), anterior (basal, mid, and apical), anterolateral (basal), and lateral (apical) were also predictive, with the basal inferoseptal longitudinal strain having the greatest estimate of feature importance amongst these longitudinal strain measures. Interestingly, transverse strain measures across these segments were also determined among the top predictive features. However, none of the radial strain measures with the exception of the mid anteroseptal segment were found to be predictive of nadir LVEF.

Figure 3.

Relative predictive importance of LV segmental strain measures. The figure summarizes feature importance, as measured by the change in mean squared error (MSE), of LV segmental strain measures at doxorubicin completion for the prediction of nadir LVEF during subsequent follow-up.

Incremental value of LV segmental strain measures for the prediction of CTRCD

We evaluated the incremental value, when added to a reference model including age, cancer therapy regimen, hypertension, and LVEF, of LV segmental strain measures that were determined to be predictive in the ML model based on change in MSE >0 (Supplementary data online, Table S12). The addition of these segmental measures to the reference model yielded an increase in the area under the curve (AUC) from 0.70 (95% CI 0.60–0.80) to 0.87 (95% CI 0.81–0.92) with a resultant gain in AUC of 0.18 (95% CI 0.08–0.27) (Figure 4). This is higher than the AUC attained with the addition of average measures of longitudinal strain (AUC = 0.72 95% CI 0.62–0.82), circumferential strain (AUC = 0.73 95% CI 0.62–0.83), transverse strain (AUC = 0.70 95% CI 0.60–0.80), or radial strain (AUC = 0.70 95% CI 0.60–0.80). Consistent findings were observed in a sensitivity analysis using an alternative definition of CTRCD based on the criteria proposed by the ASE/EACVI consensus group¹⁸ (Supplementary data online, Figure S1).

Figure 4.

Receiver operating characteristic (ROC) curves. Incremental value of LV segmental strain measures, as quantified by the area under the curve (AUC), for the prediction of CTRCD when added to a reference model (Model 1) including age, cancer therapy regimen, hypertension and LVEF at doxorubicin completion; CS, circumferential strain; LS, longitudinal strain; RS, radial strain; TS, transverse strain.

Discussion

We determined the early changes in LV segmental strain measures following doxorubicin chemotherapy and evaluated the predictive value of LV segmental measures at doxorubicin completion for subsequent LVEF declines in women with breast cancer. Our main findings are: (i) from pre-treatment to doxorubicin completion, longitudinal strain worsened across all but the apical LV segments, and circumferential strain worsened primarily in the septal LV segments, (ii) circumferential strain measures in the anterolateral and inferoseptal LV segments were the top predictive features for nadir LVEF post-doxorubicin completion; longitudinal and transverse strain measures across the inferoseptal, anterior, anterolateral, and lateral segments were also identified among the top predictive features; (iii) LV segmental strain measures provided significant incremental value for the prediction of CTRCD.

Segmental changes in LV strain measures following doxorubicin chemotherapy have been reported in prior studies.^19–25 However, these studies mostly focused on longitudinal strain, and are limited by small sample size and discrepant findings. Our study is the largest and most comprehensive study to date reporting on segmental differences in the pattern of LV strain changes following doxorubicin chemotherapy in breast cancer. We observed a significant worsening in longitudinal strain across the mid and basal LV segments, but the apical segments were largely unchanged. A non-uniform pattern of worsening in longitudinal strain has been previously described.^19–21 For instance, in a prospective cohort study of 52 women with breast cancer receiving anthracyclines, Stoodley et al.¹⁹ reported significant decreases in longitudinal strain in mid and basal LV segments but not in the apical lateral wall. A similar finding was noted in a subsequent study including 78 women with breast cancer treated with anthracyclines from this group. Significant reductions in longitudinal strain in the mid and basal LV segments were observed immediately after the completion of anthracycline chemotherapy and this persisted for 6 months after initial exposure, while the apical segments were relatively spared.²⁰ Other studies have reported a more global reduction in longitudinal strain including the apical segments.^22–25 However, most of these studies were conducted in patient populations with non-breast cancer conditions including lymphoma and sarcoma. Altogether, our study provides further evidence supporting a non-uniform pattern of worsening in longitudinal strain across LV segments following doxorubicin chemotherapy in breast cancer.

Albeit limited, there are also data suggesting a differential pattern of segmental changes in circumferential strain early after anthracycline therapy. A pattern of worsening in circumferential strain predominantly in the septal segments with no significant changes in the other LV segments has been previously described in breast cancer.¹⁹ In line with this, our findings showed reductions in circumferential strain across the septal segments, but not the anterior, lateral or inferior mid LV segments. Interestingly, there appears to be a more global worsening in radial strain, although the changes were more pronounced in the inferior and inferolateral segments, consistent with previous observations.¹⁹

Regional differences in the prognostic relevance of segmental longitudinal strain measures have been reported. For instance, a recent study by Saijo et al.,²¹ indicated that while basal longitudinal strain (i.e. the average across six basal segments) predicted CTRCD risk, the measures across the mid and apical LV regions were not associated with CTRCD in patients who received anthracyclines. We further determined the relative importance of specific segmental strain measures for the prediction of nadir LVEF post-doxorubicin completion by including both longitudinal and non-longitudinal strain using supervised ML. Circumferential strain in the anterolateral and inferoseptal segments had the greatest estimates of feature importance. Multiple longitudinal strain measures, particularly across the basal region, were also identified among the predictive features. Out of these, the basal inferoseptal segment was determined to be the most predictive. Interestingly, Mahjoob et al.²⁶ recently reported that longitudinal strain in this segment had the highest discriminatory performance for the prediction of CTRCD as compared to the other LV segments. Another important observation is that transverse strain measures in the mid inferior, apical anterior and apical lateral segments were identified among the top predictive features for nadir LVEF post-doxorubicin completion. So far, there are no data on the relevance of transverse strain for the prediction of CTRCD in patients receiving cardiotoxic cancer therapy. Our findings would suggest that transverse strain might have a potential role in cardio-oncology, although improvements in reproducibility are needed, and therefore, warrant further investigation to establish its value for CTRCD risk prediction.

Strain imaging-based cardiotoxicity risk prediction strategies have primarily focused on global strain parameters such as global longitudinal and circumferential strain. The value of segmental strain measures for the prediction of CTRCD in patients treated with anthracycline-based chemotherapy regimen is unclear. We evaluated the incremental value of LV segmental strain measures when added to a set of common predictors of CTRCD. This model had an AUC of 0.70, which is comparable to other published clinical risk prediction models.^5–7 The addition of the predictive segmental strain measures identified using supervised ML significantly increased the AUC to 0.87. This gain is higher when compared with the incremental value of average longitudinal, circumferential, transverse, or radial strain measures. This might suggest that a more comprehensive utilization of LV segmental strain beyond the global measures could further enhance the value of strain imaging for the prediction of CTRCD.

Besides external validation in independent cohorts, a greater understanding of the mechanisms underlying the differences in the pattern of worsening and predictive value for subsequent cardiac dysfunction is critical before the widespread clinical implementation of segmental strain can be considered. Currently, these pathophysiologic mechanisms are not well characterized. We postulate that disparities in local LV geometry and its interaction with wall stress could be among the possible mechanistic explanations. Based on the Laplace law, LV curvature radius is directly related to wall stress, and therefore, given the greater radius at the base of the LV, local wall stress is higher in the basal and mid regions, and lower in the apical region.²⁷ Consequently, myocardial fibres in the basal and mid regions, particularly the longitudinal fibres, must work against a higher load even under normal conditions.²⁸ Interestingly, significant early increases in aortic stiffness (an indirect marker of LV afterload) and LV wall stress have been reported following doxorubicin chemotherapy.^12,29 We hypothesize that this increased afterload might result in a greater susceptibility of the myocardial fibres in the basal and mid regions to adverse doxorubicin-induced pathophysiologic changes such as oxidative and nitrosative stress, and endothelial dysfunction when compared with the fibres in the apical region. This might partly explain the non-uniform pattern of change in longitudinal strain observed in our study. Myocardial fibre orientation and its relationship with stress-induced functional changes should also be considered as a possible mechanistic explanation. Multiple studies indicate that longitudinal fibres (particularly subendocardial fibres) are affected early following ischaemia or hemodynamic overload, and hence longitudinal function worsens early. Meanwhile, circumferential fibres are less prone to damage, and indeed, there is strong evidence suggesting that circumferential function is preserved even when longitudinal function is significantly impaired helping to maintain global systolic function.^28,30,31 This might help explain our observation that circumferential strain, but not longitudinal strain, is largely preserved across most mid-LV segments. In addition, it might partly explain the finding that circumferential strain measures had the greatest feature importance for the prediction of nadir LVEF during subsequent follow-up. Doxorubicin has also been recently shown to induce myocardial fibrosis at an early stage, and the possibility of inherent regional differences in the pattern of fibrotic changes and remodelling across the LV wall need further consideration.³² Altogether, our findings suggest that more basic, mechanistic studies are warranted to better understand cardiac remodelling following doxorubicin chemotherapy.

Several limitations need further consideration. Although we utilized core lab echocardiographic quantitation to maximize reproducibility, variability of segmental strain measures, particularly the radial and transverse dimensions, is somewhat higher than typically reported for global strain measures. There were also differences in analysability between different imaging planes although the image quality, reliability, and predictive associations were generally comparable across different views. Due to limited image analysability, our study lacks apical three-chamber data, and the potential importance of segmental strain measures derived from this view cannot be ascertained. Although our study is the largest to date to examine changes in segmental strain in cardio-oncology, our sample size is relatively small for ML analysis. In addition, our study lacks external validation of findings in an independent cohort of women with breast cancer. Other possible limitations of ML methods such as model overfitting also need consideration. Although we performed 10-fold cross-validation for model training to reduce overfitting, this does not completely eliminate the possibility of model overfitting. As a result of these limitations, we consider our results to be hypothesis-generating. Larger studies are needed to validate our findings prior to implementation of segmental strain assessment in clinical practice. There were also a limited number of CTRCD events and we were unable to develop random forest classifier models for this categorical outcome. Besides doxorubicin, patients also received non-anthracycline chemotherapy agents such as cyclophosphamide and taxanes that have been shown to be potentially associated with increased risk of cardiomyopathy in previous studies.³³ The potential impact of such chemotherapy agents need further consideration in the interpretation of our findings.

In conclusion, in this prospective cohort study of women with breast cancer who received doxorubicin chemotherapy, we determined segmental differences in the patterns of worsening of LV strain, particularly for longitudinal and circumferential strain. We determined that measures at the completion of doxorubicin of circumferential strain in the mid anterolateral and inferoseptal segments and, longitudinal and transverse strain measures across the inferoseptal, anterior and lateral segments are among the most predictive features for nadir LVEF post-doxorubicin completion. Our findings also suggest that segmental strain measures may provide significant incremental value for the prediction of CTRCD. Our study serves as a proof-of-principle, generating important hypotheses and providing motivation for future studies aimed at developing highly accurate imaging-based risk prediction models using ML.

Supplementary data

Supplementary data are available at European Heart Journal - Cardiovascular Imaging online.

Funding

This work was supported by the American Heart Association Institute for Precision Cardiovascular Medicine Uncovering New Patterns in Cardiovascular Disease, NHLBI R01-HL118018, American Cancer Society Institutional Research 78-002-30, NHLBI K23-HL095661. The funding organizations had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; or decision to submit the manuscript for publication.

Conflict of interest: none declared.