Cardiac Reasoning Template (CRT)

Updated 20 January 2026

Cardiac Reasoning Template (CRT) is a framework that formalizes multi-stage, uncertainty-aware decision pipelines for cardiac diagnosis and risk prediction.
It modularizes workflows from data ingestion to model deployment, integrating ensemble methods, uncertainty quantification, and logical reasoning for clinical decision support.
CRTs combine techniques from predictive modeling, large language models, and quantitative imaging to improve interpretability, efficiency, and actionable outcomes in cardiac care.

A Cardiac Reasoning Template (CRT) formalizes standardized, transparent, and actionable pipelines for cardiac decision support, encompassing multi-stage modeling, uncertainty quantification, knowledge-guided logical flows, and explicit reasoning steps. CRTs capture and modularize the workflow underlying clinical diagnosis, risk prediction, and interpretability in cardiac medicine, from multimodal data ingestion to model deployment and clinical application.

1. Multi-Stage and Uncertainty-Aware CRT Pipelines

CRT frameworks are exemplified by multi-stage decision logic, particularly in cardiac resynchronization therapy (CRT) response prediction for heart failure patients (Larsen et al., 2023). The canonical architecture initiates with a Stage 1 ensemble, typically elastic-net logistic regression models trained on demographic, clinical, and 12-lead ECG features (e.g., age, gender, comorbidities, QRS duration, LBBB presence). Each ensemble member outputs a response probability $p_m$ ; the mean and standard deviation ( $\hat p, \sigma$ ) across $M$ ensemble members estimate the model's predictive certainty.

If the Stage 1 prediction meets certainty thresholds (e.g., $\sigma \leq \tau$ and $|\hat p - 0.5| \geq \delta$ ), the result is accepted. Otherwise, the pipeline “cascades” to Stage 2, acquiring additional features (such as resting gated SPECT MPI variables—LVEF, volumes, scar burden, shape/dyssynchrony indices), which are fed into a second, augmented ensemble (Ensemble 2). This logic significantly reduces unnecessary imaging (SPECT used in only 52.7% of cases) without sacrificing AUC, accuracy, sensitivity, or specificity relative to “always-use-all-modalities” models.

Model/Stage	AUC	Accuracy	Sensitivity	Specificity
Ensemble 1 (Clinical+ECG)	0.70 ± 0.08	0.64 ± 0.07	0.61 ± 0.09	0.67 ± 0.13
Ensemble 2 (+SPECT MPI)	0.77 ± 0.08	0.69 ± 0.11	0.72 ± 0.14	0.65 ± 0.15
Multi-Stage CRT Model	0.75 ± 0.10	0.71 ± 0.11	0.70 ± 0.13	0.72 ± 0.20
Guideline Criteria	—	0.53	0.75	0.26

2. CRTs for Knowledge-Driven Clinical Reasoning and LLM Training

CRTs are also formalized as disease-specific templates for modular, stepwise diagnostic reasoning in complex cardiac diseases, as developed in the context of multimodal LLMs (MLLMs) for interpretable echocardiographic diagnosis (Qin et al., 13 Jan 2026). Each CRT instance is a triple:

$T = \{$ $T = {$ T_name, T_meta, T_reason $\}$ $}$ , where
- T_name: disease/task label (e.g., Hypertrophic Cardiomyopathy)
- T_meta: fielded metadata (tags, procedural descriptions, required imaging/measures)
- T_reason: ordered list of $m$ logical steps (e.g., “Evaluate septal hypertrophy on PLAX”, “Assess diastolic filling on Doppler”).

During MLLM training, the CRT is retrieved by meta-fields, prepended to the prompt, and acts as a high-level skeleton. The model fills in detailed chains of thought, which are quantitatively rewarded for step count (Procedural Quantity Reward, PQtR), step relevance (Procedural Quality Reward, PQlR), and semantic grounding in echocardiographic content (Echocardiographic Semantic Reward, ESR). This template-informed process resulted in a 48% improvement in diagnostic precision for multiview echo and 93.3% clinician agreement with the generated reasoning (Qin et al., 13 Jan 2026).

3. Hierarchical and Multimodal Reasoning in Survival Analysis

In postoperative cardiovascular risk prediction, the CRT has been implemented as a hierarchical two-stage pipeline combining large language/vision models and Cox survival analysis (Rui et al., 25 May 2025).

Stage 1: Hierarchical reasoning trajectories ( $Z$ ) are constructed using evidence-augmented, self-refinement LLM loops over the radiological findings, clinical text, and image data. Iterative review, correction, and endpoint consistency checks lead to a refined multi-part reasoning trace (diagnosis, complications, follow-up).
Stage 2: Quantitative image features (MRI embeddings) and encoded reasoning traces are fused via attention mechanisms into a unified Cox-based survival prediction.

Interpretability is explicit: the CRT exposes token-level reasoning subchains for every diagnostic/complication/follow-up inference, and ablation confirms their necessity for optimal prognostic performance (C-index up to 0.83-0.84, outperforming unimodal or non-reasoning baselines).

4. CRTs as Modular Decision-Support in CRT and Imaging Workflows

CRTs have been leveraged for explicit, template-driven decision support in CRT eligibility and outcome prediction (Fernandes et al., 2022 Larsena et al., 2023), with the following structure:

Patient eligibility assessment (demographic, clinical, and ECG screening)
Multi-modal input acquisition (clinical, ECG, SPECT/GMPS, CMR)
Feature derivation (e.g., $\Delta$ LVEF, phase/dyssynchrony, scar burden)
ML model application (e.g., PAM, Naive Bayes, neural networks, DL with image fusion)
Output: class assignment (“responder”, “super-responder”), probability score, key variable explanations

CRT-based models systematically outperform guideline algorithms (AUC up to 0.86-0.87, specificity gains 0.22 vs. 0.75), with polarmap-driven deep learning architectures further improving predictive metrics (AUC 0.83, accuracy 0.73) (Larsena et al., 2023). CRTs thus structure the entire workflow from inclusion/exclusion to actionable, interpretable prediction.

5. CRTs in Quantitative Cardiac Imaging and Signal Processing

In quantitative cardiac MRI, CRTs are instantiated as complete motion-correction and mapping pipelines leveraging deep-learning-based groupwise registration (Zhang et al., 2024). The “PCA-Relax” CRT consists of:

U-Net-based deformable registration operating groupwise (no explicit template selection)
PCA loss enforces maximal alignment across image times
Relaxometry loss regularizes registration by enforcing signal-model conformity voxelwise
Additional regularizers include displacement smoothness and cyclic-consistency

This CRT results in lower myocardial fitting SD and robust mapping across contrast variation and subject-specific motion, outperforming pairwise and template-based methods across multiple evaluation strategies.

6. Template Construction and Implementation Considerations

CRTs are explicitly modular and extensible (multi-stage logic, reward-based RL, hierarchical self-refinement, attention-based fusion), promoting broad adaptation:

Rule-logic enables selective downstream test acquisition, minimizing unnecessary cost and burden (e.g., SPECT or CMR gating contingent on model uncertainty).
CRT meta-fields enable search, retrieval, and in-context guidance for both algorithmic models and LLM agents.
Extensible to new modalities (e.g., speckle-tracking echo, T1/T2 mapping, multi-omics) or incremental biomarkers.
Embedded mathematical notation provides precise thresholds, quantitative rules, and state/logical transitions for direct real-world deployment.

7. Clinical Impact and Future Directions

The CRT paradigm advances interpretability, efficiency, and transparency in cardiology AI systems by operationalizing expert workflows into auditable, modular pipelines. In multi-stage CRT decision support, expert-level diagnostic reasoning, and quantitative imaging registration, CRTs formalize not just the “what” but the “how” of cardiac inference—demarcating trinities of data input, logical processing, and actionable output. As additional modalities and endpoints are incorporated, CRTs are anticipated to serve as the foundational “operating system” for real-world, explainable, and equitable cardiac AI.

References: