Expanded adaptation of an artificial intelligence model for predicting chemotherapy-induced cardiotoxicity using baseline electrocardiograms

Introduction

While effective as chemotherapeutics, anthracyclines can cause cancer therapy-related cardiac dysfunction (CTRCD), which adversely affects the prognosis of patients with malignancies1–5. Despite guideline recommendations6–9, repeated echocardiograms are rarely performed10 with delayed diagnosis of CTRCD leading to unrecoverable cardiac dysfunction11. Recently, artificial intelligence (AI) was shown to be capable of detecting reduced left ventricular ejection fraction (LVEF) solely from electrocardiogram (ECG)12. Furthermore, this model was predictive of a future decrease in LVEF. Therefore, we hypothesized that an AI model detecting reduced LVEF (AI-EF model) could predict CTRCD from ECGs.

Purpose

To assess whether the AI-EF model could detect patients at a high risk of CTRCD by analyzing ECGs taken immediately prior to the initiation of cardiotoxic chemotherapy.

Methods

Among patients who received chemotherapy with a regimen including anthracyclines in two institutions between June 1st, 2015 and October 1st, 2020, those who underwent both an ECG and echocardiogram ≤90 days prior to initial treatment were selected. The ECGs were analyzed by the AI-EF model and patients were divided into two groups according to the scores from the model. CTRCD was defined as LVEF <53% and ≥10% decrease in LVEF from the baseline at any time after the start of chemotherapy13. The cumulative incidence of CTRCD was compared for the two groups using Kaplan-Meier curves, log-rank test, a univariate Cox proportional hazard model, and a multivariable Cox proportional hazard model adjusting for known risk factors for CTRCD. Finally, a prediction model for CTRCD using readily available clinical variables with the AI-EF score was compared with the model using the same variables without the AI-EF score.

Results

1,158 patients were included in this study. 99 of them developed CTRCD during follow-up. The AI-EF model displayed excellent risk stratification of developing CTRCD: while 7.1% in the low AI-EF score group developed CTRCD, 12.9% of the patients in the high AI-EF score group developed CTRCD (hazard ratio (HR), 2.14; 95% confidence interval (CI), 1.43–3.19; log-rank p<0.001; Figure 1). This finding was robust across subgroups such as cancer types, the initial dose of anthracycline and baseline LVEF, and consistent after adjusting for multiple risk factors (adjusted HR, 2.10; 95% CI, 1.37–3.22; p<0.001; Figure 2). Furthermore, the addition of the AI-EF score significantly improved the accuracy of predicting CTRCD compared to clinical features alone (time-dependent area under the received operating curve (AUROC) for 2 years, 77.1; 95% CI, 71.8–82.3 for the model with AI-EF score and AUROC 73.9; 95% CI, 69.0–80.1 for the model without AI-EF score; p=0.02).

Conclusion

The AI-EF model, by utilizing baseline ECG, could stratify patients according to the risk of CTRCD and robustly augmented CTRCD prediction.

Funding Acknowledgement

Type of funding sources: Private grant(s) and/or Sponsorship. Main funding source(s): American Heart AssociationVerily

Fast-convergent iterative scheme for filtering velocity signals and finding Kolmogorov scales

The present fundamental knowledge of fluid turbulence has been established primarily from hot- and cold-wire measurements. Unfortunately, however, these measurements necessarily suffer from contamination by noise since no certain method has previously been available to optimally filter noise from the measured signals. This limitation has impeded our progress of understanding turbulence profoundly. We address this limitation by presenting a simple, fast-convergent iterative scheme to digitally filter signals optimally and find Kolmogorov scales definitely. The great efficacy of the scheme is demonstrated by its application to the instantaneous velocity measured in a turbulent jet.

X-ray structure of the human hyperplastic discs protein: an ortholog of the C-terminal domain of poly(A)-binding protein

The poly(A)-binding protein (PABP) recognizes the 3' mRNA poly(A) tail and plays an essential role in eukaryotic translation initiation and mRNA stabilization/degradation. PABP is a modular protein, with four N-terminal RNA-binding domains and an extensive C terminus. The C-terminal region of PABP is essential for normal growth in yeast and has been implicated in mediating PABP homo-oligomerization and protein-protein interactions. A small, proteolytically stable, highly conserved domain has been identified within this C-terminal segment. Remarkably, this domain is also present in the hyperplastic discs protein (HYD) family of ubiquitin ligases. To better understand the function of this conserved region, an x-ray structure of the PABP-like segment of the human HYD protein has been determined at 1.04-A resolution. The conserved domain adopts a novel fold resembling a right-handed supercoil of four alpha-helices. Sequence profile searches and comparative protein structure modeling identified a small ORF from the Arabidopsis thaliana genome that encodes a structurally similar but distantly related PABP/HYD domain. Phylogenetic analysis of the experimentally determined (HYD) and homology modeled (PABP) protein surfaces revealed a conserved feature that may be responsible for binding to a PABP interacting protein, Paip1, and other shared interaction partners.

Translational repression by a novel partner of human poly(A) binding protein, Paip2

The eukaryotic mRNA 3' poly(A) tail acts synergistically with the 5' cap structure to enhance translation. This effect is mediated by a bridging complex, composed of the poly(A) binding protein (PABP), eIF4G, and the cap binding protein, eIF4E. PABP-interacting protein 1 (Paip1) is another factor that interacts with PABP to coactivate translation. Here, we describe a novel human PABP-interacting protein (Paip2), which acts as a repressor of translation both in vitro and in vivo. Paip2 preferentially inhibits translation of a poly(A)-containing mRNA, but has no effect on the translation of hepatitis C virus mRNA, which is cap- and eIF4G-independent. Paip2 decreases the affinity of PABP for polyadenylate RNA, and disrupts the repeating structure of poly(A) ribonucleoprotein. Furthermore, Paip2 competes with Paip1 for PABP binding. Thus, Paip2 inhibits translation by interdicting PABP function.

Recognition of polyadenylate RNA by the poly(A)-binding protein

The cocrystal structure of human poly(A)-binding protein (PABP) has been determined at 2.6 A resolution. PABP recognizes the 3' mRNA poly(A) tail and plays critical roles in eukaryotic translation initiation and mRNA stabilization/degradation. The minimal PABP used in this study consists of the N-terminal two RRM-type RNA-binding domains connected by a short linker (RRM1/2). These two RRMs form a continuous RNA-binding trough, lined by an antiparallel beta sheet backed by four alpha helices. The polyadenylate RNA adopts an extended conformation running the length of the molecular trough. Adenine recognition is primarily mediated by contacts with conserved residues found in the RNP motifs of the two RRMs. The convex dorsum of RRM1/2 displays a phylogenetically conserved hydrophobic/acidic portion, which may interact with translation initiation factors and regulatory proteins.