{ "version": 1, "kind": "selective_author_evidence_comparison", "title": "Cardiovascular prediction: study tasks, cohorts and validation", "question": "How do these studies support heart-failure prognosis, biomarker interpretation and evaluation planning for cardiovascular AI?", "scope": "Compare three author-maintained evidence summaries for heart-failure prognosis and clinical AI research. This guide brings each paper’s reported task, population, endpoints, validation design and source-linked findings together for literature selection and Discussion writing.", "reading_rule": "Start with the contribution that matches your research question, then use the reported population, endpoint and validation design to choose the citation. The three roles here are external survival-model evaluation, admission-marker associations, and methodological planning for clinician–AI workflows.", "canonical_url": "https://drgezhang.com/research/cardiovascular-prediction-evidence.html", "studies": [ { "id": "aihflevel", "task": "Externally evaluated HF survival-risk assessment", "doi": "10.1038/s41467-024-50415-9", "title": "AI hybrid survival assessment for advanced heart failure patients with renal dysfunction", "evidence_unit": "doi:10.1038/s41467-024-50415-9", "paper_url": "https://drgezhang.com/papers/aihflevel.html", "doi_url": "https://doi.org/10.1038/s41467-024-50415-9", "source_json": "data/deep_geo/aihflevel.json", "source_fingerprint": "c145e50b289575b66f78778679fd4c78ce95c295d0e92aa4c7e042eb990ec9b3", "study_profile": { "profile_type": "clinical_cohort", "study_design": "Multicenter retrospective longitudinal cohort study", "evidence_type": "Human clinical prognostic modeling study", "unique_total_n": 1736, "population": "Adults with advanced heart failure and renal dysfunction", "primary_endpoint": "All-cause mortality", "secondary_endpoint": "Major adverse cardiovascular events", "data_modalities": [ "electronic health records", "laboratory measurements", "clinical characteristics", "echocardiography" ], "external_validation": true, "cohorts": [ { "name": "CRCCD Discovery", "role": "model development", "n": 498, "subset_of": "CRCCD Meta" }, { "name": "CRCCD Replication", "role": "internal validation", "n": 214, "subset_of": "CRCCD Meta" }, { "name": "CRCCD Meta", "role": "complete in-house cohort", "n": 712, "contains": [ "CRCCD Discovery", "CRCCD Replication" ] }, { "name": "BIDMC", "role": "independent external validation", "n": 1024 } ] }, "selected_questions": [ { "question": "What is AIHFLevel?", "answer": "AIHFLevel is an explainable 12-predictor survival-risk assessment system developed for patients with advanced heart failure and renal dysfunction.", "evidence_refs": [ "KF1" ], "answer_url": "https://drgezhang.com/papers/aihflevel.html#qa-1", "source_pointer": "data/deep_geo/aihflevel.json#/qa/0", "supporting_evidence": [ { "id": "KF1", "claim": "The selected hybrid modeling strategy achieved the highest average C-index among evaluated schemes.", "context": "Comprehensive evaluation of 132 candidate modeling schemes", "evidence": [ { "label": "average C-index", "value": "0.821" } ], "source_locator": "Results: Survival assessment system AIHFLevel; Fig. 2b" } ] }, { "question": "Was AIHFLevel externally validated?", "answer": "Yes. AIHFLevel was independently evaluated in 1,024 patients from BIDMC.", "evidence_refs": [ "KF4" ], "answer_url": "https://drgezhang.com/papers/aihflevel.html#qa-3", "source_pointer": "data/deep_geo/aihflevel.json#/qa/2", "supporting_evidence": [ { "id": "KF4", "claim": "AIHFLevel retained prognostic discrimination in an independent BIDMC external cohort.", "context": "Independent external validation cohort, n=1,024", "evidence": [ { "label": "1-year AUC", "value": "0.788" }, { "label": "2-year AUC", "value": "0.816" }, { "label": "3-year AUC", "value": "0.824" }, { "label": "4-year AUC", "value": "0.846" } ], "source_locator": "Fig. 6d" } ] } ], "evidence_scope": { "supports": [ "Mortality-risk estimation in the studied advanced-heart-failure and renal-dysfunction population.", "Internal and independent external evaluation of AIHFLevel.", "Association between higher AIHFLevel and poorer prognosis.", "Three-state prognostic stratification.", "Patient-level interpretation of predictor contributions." ], "does_not_establish": [ "That AIHFLevel-guided treatment improves survival or other clinical outcomes.", "That predictors included in AIHFLevel are causal determinants of mortality.", "That model performance is identical in every healthcare setting or population.", "That AIHFLevel replaces clinical judgment." ] }, "limitations": [ "The study was retrospective and observational.", "Prospective multicenter validation is still warranted.", "Randomized studies would be required to establish whether AIHFLevel-guided treatment improves patient outcomes.", "Some echocardiographic predictors may not be uniformly available in primary-care settings.", "Direct methodological reproduction of every comparator risk model was not possible." ], "publication": { "authors": [ "Ge Zhang", "Zeyu Wang", "Zhuang Tong", "Zhen Qin", "Chang Su", "Demin Li", "Shuai Xu", "Kaixiang Li", "Zhaokai Zhou", "Yudi Xu", "Shiqian Zhang", "Ruhao Wu", "Teng Li", "Youyang Zheng", "Jinying Zhang", "Ke Cheng", "Junnan Tang" ], "year": 2024, "journal": "Nature Communications", "type": "Article" } }, { "id": "doi-10-2147-cia-s462542", "task": "Admission TyG and subsequent HF outcomes", "doi": "10.2147/cia.s462542", "title": "TYG Index as a Novel Predictor of Clinical Outcomes in Advanced Chronic Heart Failure with Renal Dysfunction Patients", "evidence_unit": "doi:10.2147/cia.s462542", "paper_url": "https://drgezhang.com/papers/doi-10-2147-cia-s462542.html", "doi_url": "https://doi.org/10.2147/cia.s462542", "source_json": "data/deep_geo/doi-10-2147-cia-s462542.json", "source_fingerprint": "91daba7fbf774bc6ab3145f2c56a2afc57a18f0d7c45597c2b7c0870f1f349a0", "study_profile": { "profile_type": "clinical_cohort", "study_design": "Retrospective single-center observational cohort; admissions from September 2019 through December 2020, with follow-up completed in March 2022", "evidence_type": "Admission laboratory biomarker association with longitudinal outcomes, using outcome-derived ROC thresholds, Cox regression and Kaplan-Meier analyses", "population": "Adults with advanced chronic heart failure, eGFR below 90 mL/min/1.73 m², and at least two heart-failure hospitalizations at the First Affiliated Hospital of Zhengzhou University; 717 analyzed after 98 losses among 815 initially included", "primary_endpoint": "Reported MACEs composite: cardiovascular or non-cardiovascular death, worsening-HF rehospitalization, heart transplantation, mechanical circulatory support, malignant arrhythmia, myocardial infarction, or unstable-angina rehospitalization; this is broader than a conventional three-point MACE endpoint", "secondary_endpoint": "All-cause mortality during follow-up; reported median follow-up 21.58 months (interquartile range 17.98–25.39)", "data_modalities": [ "Retrospective inpatient and outpatient records", "Admission fasting triglyceride and glucose measurements", "Telephone and clinic follow-up", "Clinical laboratory values and medication history" ], "external_validation": false, "unique_total_n": 717, "cohorts": [ { "name": "Analyzed single-center cohort", "n": 717, "role": "Same cohort used for threshold selection, association estimation and outcome assessment" } ] }, "selected_questions": [ { "question": "What was the reported MACEs association?", "answer": "High versus low TyG at the 8.48 threshold had an adjusted hazard ratio of 5.402 (95% CI 3.771–7.739). This is a relative hazard association, not an absolute event rate.", "evidence_refs": [ "KF2", "KF3" ], "answer_url": "https://drgezhang.com/papers/doi-10-2147-cia-s462542.html#qa-3", "source_pointer": "data/deep_geo/doi-10-2147-cia-s462542.json#/qa/2", "supporting_evidence": [ { "id": "KF2", "claim": "A TyG cutoff of 8.48 separated groups with different composite-event risks, with moderate discrimination in the source cohort.", "context": "The endpoint-specific cutoff and ROC performance were derived and evaluated in the same cohort; no independent external validation was reported.", "evidence": [ { "label": "MACEs cutoff", "value": "8.48" }, { "label": "ROC AUC", "value": "0.759" }, { "label": "Sensitivity", "value": "0.836" }, { "label": "Specificity", "value": "0.588" }, { "label": "Low TyG group", "value": "318 patients, TyG <8.48" }, { "label": "High TyG group", "value": "399 patients, TyG ≥8.48" } ], "source_locator": "PMC11068055: Results — Baseline Characteristics; Figure 2; Table 1 (Clinical and Laboratory Characteristics According to the TyG (MACEs), anchor t0001)." }, { "id": "KF3", "claim": "Higher TyG was associated with a higher adjusted hazard of the reported composite adverse-event endpoint.", "context": "The comparison is the high-TyG group versus the low-TyG group in an observational multivariable Cox model. The hazard ratio is not an absolute event probability or a treatment effect.", "evidence": [ { "label": "Adjusted MACEs hazard ratio", "value": "5.402" }, { "label": "95% confidence interval", "value": "3.771–7.739" }, { "label": "Reported significance", "value": "P <0.001 in Results; table rounds P to 0" } ], "source_locator": "PMC11068055: Table 5 — Multivariate Cox Regression Analysis Results for MACEs, TyG row (anchor t0005); Results — Relationship Between TyG and MACEs. The narrative cross-reference points to Table 4, but the verified result is in Table 5." } ] }, { "question": "Who was included in the analyzed cohort?", "answer": "The analysis included 717 adults with advanced chronic HF, study-defined renal dysfunction and repeated HF hospitalization at one center. The 815 initially included patients also contained 98 who were lost to follow-up.", "evidence_refs": [ "KF1" ], "answer_url": "https://drgezhang.com/papers/doi-10-2147-cia-s462542.html#qa-1", "source_pointer": "data/deep_geo/doi-10-2147-cia-s462542.json#/qa/0", "supporting_evidence": [ { "id": "KF1", "claim": "The analysis comprised 717 patients after follow-up losses, with outcome follow-up of about two years.", "context": "This is an observational advanced-heart-failure population selected for renal dysfunction and repeated HF hospitalization; the initial included population is not the analyzed sample size.", "evidence": [ { "label": "Initially included", "value": "815 patients" }, { "label": "Lost to follow-up", "value": "98 patients (12.0%)" }, { "label": "Analyzed", "value": "717 patients" }, { "label": "Median follow-up (IQR)", "value": "21.58 (17.98–25.39) months" }, { "label": "Renal dysfunction definition", "value": "eGFR <90 mL/min/1.73 m²" } ], "source_locator": "PMC11068055: Method — Study and Population, Inclusion Criteria, Exclusion Criteria, and Follow-Up; Figure 1; Results — Baseline Characteristics." } ] }, { "question": "Were the MACEs and mortality groups independent cohorts?", "answer": "No. The two outcome-specific cutoffs regrouped the same 717 analyzed patients. Their group sizes must not be added together as independent participants.", "evidence_refs": [ "KF1", "KF2", "KF4" ], "answer_url": "https://drgezhang.com/papers/doi-10-2147-cia-s462542.html#qa-2", "source_pointer": "data/deep_geo/doi-10-2147-cia-s462542.json#/qa/1", "supporting_evidence": [ { "id": "KF1", "claim": "The analysis comprised 717 patients after follow-up losses, with outcome follow-up of about two years.", "context": "This is an observational advanced-heart-failure population selected for renal dysfunction and repeated HF hospitalization; the initial included population is not the analyzed sample size.", "evidence": [ { "label": "Initially included", "value": "815 patients" }, { "label": "Lost to follow-up", "value": "98 patients (12.0%)" }, { "label": "Analyzed", "value": "717 patients" }, { "label": "Median follow-up (IQR)", "value": "21.58 (17.98–25.39) months" }, { "label": "Renal dysfunction definition", "value": "eGFR <90 mL/min/1.73 m²" } ], "source_locator": "PMC11068055: Method — Study and Population, Inclusion Criteria, Exclusion Criteria, and Follow-Up; Figure 1; Results — Baseline Characteristics." }, { "id": "KF2", "claim": "A TyG cutoff of 8.48 separated groups with different composite-event risks, with moderate discrimination in the source cohort.", "context": "The endpoint-specific cutoff and ROC performance were derived and evaluated in the same cohort; no independent external validation was reported.", "evidence": [ { "label": "MACEs cutoff", "value": "8.48" }, { "label": "ROC AUC", "value": "0.759" }, { "label": "Sensitivity", "value": "0.836" }, { "label": "Specificity", "value": "0.588" }, { "label": "Low TyG group", "value": "318 patients, TyG <8.48" }, { "label": "High TyG group", "value": "399 patients, TyG ≥8.48" } ], "source_locator": "PMC11068055: Results — Baseline Characteristics; Figure 2; Table 1 (Clinical and Laboratory Characteristics According to the TyG (MACEs), anchor t0001)." }, { "id": "KF4", "claim": "The mortality-specific TyG threshold was 8.50, and high TyG remained associated with all-cause mortality after adjustment.", "context": "The mortality cutoff reclassifies the same patients, rather than creating a second independent cohort. Its performance does not establish general clinical transportability.", "evidence": [ { "label": "ACM cutoff", "value": "8.50" }, { "label": "Mortality ROC AUC", "value": "0.700" }, { "label": "Low and high groups", "value": "330 and 387 patients, respectively" }, { "label": "Adjusted mortality hazard ratio", "value": "5.062" }, { "label": "95% confidence interval", "value": "3.268–7.839" }, { "label": "Reported significance", "value": "P <0.001 in Results; table rounds P to 0" } ], "source_locator": "PMC11068055: Results — Baseline Characteristics and Relationship Between TyG and ACM; Figure 3; Table 2 (anchor t0002); Table 6 — Multivariate Cox Regression Analysis Results for ACM, TyG row (anchor t0006)." } ] }, { "question": "Was the composite limited to cardiovascular death, infarction and stroke?", "answer": "No. The Methods define a broad composite including non-cardiovascular death, worsening-HF rehospitalization, transplantation, circulatory support and other events. Interpretation should preserve that study-specific definition.", "evidence_refs": [ "KF1", "KF3" ], "answer_url": "https://drgezhang.com/papers/doi-10-2147-cia-s462542.html#qa-4", "source_pointer": "data/deep_geo/doi-10-2147-cia-s462542.json#/qa/3", "supporting_evidence": [ { "id": "KF1", "claim": "The analysis comprised 717 patients after follow-up losses, with outcome follow-up of about two years.", "context": "This is an observational advanced-heart-failure population selected for renal dysfunction and repeated HF hospitalization; the initial included population is not the analyzed sample size.", "evidence": [ { "label": "Initially included", "value": "815 patients" }, { "label": "Lost to follow-up", "value": "98 patients (12.0%)" }, { "label": "Analyzed", "value": "717 patients" }, { "label": "Median follow-up (IQR)", "value": "21.58 (17.98–25.39) months" }, { "label": "Renal dysfunction definition", "value": "eGFR <90 mL/min/1.73 m²" } ], "source_locator": "PMC11068055: Method — Study and Population, Inclusion Criteria, Exclusion Criteria, and Follow-Up; Figure 1; Results — Baseline Characteristics." }, { "id": "KF3", "claim": "Higher TyG was associated with a higher adjusted hazard of the reported composite adverse-event endpoint.", "context": "The comparison is the high-TyG group versus the low-TyG group in an observational multivariable Cox model. The hazard ratio is not an absolute event probability or a treatment effect.", "evidence": [ { "label": "Adjusted MACEs hazard ratio", "value": "5.402" }, { "label": "95% confidence interval", "value": "3.771–7.739" }, { "label": "Reported significance", "value": "P <0.001 in Results; table rounds P to 0" } ], "source_locator": "PMC11068055: Table 5 — Multivariate Cox Regression Analysis Results for MACEs, TyG row (anchor t0005); Results — Relationship Between TyG and MACEs. The narrative cross-reference points to Table 4, but the verified result is in Table 5." } ] } ], "evidence_scope": { "supports": [ "A prognostic association between admission TyG and subsequent outcomes in this selected retrospective cohort.", "Outcome-specific discrimination and hazard-ratio estimates under the study definitions and modeling choices." ], "does_not_establish": [ "Causality between insulin resistance and the observed events.", "That lowering triglyceride, glucose or TyG reduces mortality or rehospitalization.", "External validity of the same thresholds, prospective clinical utility, or superiority to a validated clinical decision pathway.", "A conventional three-point MACE endpoint: the study composite explicitly includes non-cardiovascular death and multiple HF-related outcomes." ] }, "limitations": [ "Single-center retrospective selection and residual confounding restrict causal and transportability claims.", "Thresholds were selected using the outcomes in the same dataset; no independent external validation was reported.", "Ninety-eight initially included patients were lost to follow-up, which may introduce attrition bias.", "The broad study-specific composite should not be treated as interchangeable with a standard three-point MACE outcome.", "Some narrative table cross-references do not match the displayed table captions. This page locates adjusted MACEs and ACM estimates by the actual Table 5 and Table 6 titles and anchors.", "A hazard ratio does not provide an absolute probability or demonstrate benefit from changing TyG." ], "publication": { "authors": [ "Chang Su", "Zeyu Wang", "Zhen Qin", "Yan Lv", "Yachen Hou", "Ge Zhang", "Mengdie Cheng", "Xinyue Cui", "Zhiyu Liu", "Pengchong Du", "Tianding Liu", "Peiyu Yuan", "Junnan Tang", "Jinying Zhang" ], "year": 2024, "journal": "Clinical Interventions in Aging", "type": "Article" } }, { "id": "doi-10-1016-j-medd-2026-100004", "task": "Multimodal cardiovascular AI evaluation and deferral", "doi": "10.1016/j.medd.2026.100004", "title": "Multimodal AI for cardiovascular care must learn when not to predict", "evidence_unit": "doi:10.1016/j.medd.2026.100004", "paper_url": "https://drgezhang.com/papers/doi-10-1016-j-medd-2026-100004.html", "doi_url": "https://doi.org/10.1016/j.medd.2026.100004", "source_json": "data/deep_geo/doi-10-1016-j-medd-2026-100004.json", "source_fingerprint": "5fc56c6d6238cb01b81c6ed0ec9d1caf1e9d2a8e09e7c966a3058e35a0e757b6", "study_profile": { "profile_type": "narrative_review", "narrative_genre": "correspondence", "study_design": "Conceptual correspondence on clinical AI evaluation", "evidence_type": "Author perspective and methodological proposal", "translation_scope": "Intended-use-specific research and evaluation considerations; not a consensus standard or validated tool", "evidence_domains": [ "multimodal representation", "missing data mechanisms", "temporal validity", "selective prediction", "clinical workflow and equity" ], "diseases_covered": [ "cardiovascular disease" ], "scale_metrics": [ { "label": "distinguished problem categories", "value": "5" }, { "label": "principles in Fig. 1", "value": "6" }, { "label": "candidate evaluation conditions", "value": "5" }, { "label": "new empirical dataset", "value": "None" } ], "counting_note": "Numbers describe the conceptual framework, not patients, external cohorts, performance estimates or trial evidence." }, "selected_questions": [ { "question": "What problem does this correspondence address?", "answer": "It addresses when multimodal cardiovascular AI has enough current and clinically relevant evidence to predict, and when it should defer.", "evidence_refs": [ "KF1", "KF2" ], "answer_url": "https://drgezhang.com/papers/doi-10-1016-j-medd-2026-100004.html#qa-1", "source_pointer": "data/deep_geo/doi-10-1016-j-medd-2026-100004.json#/qa/0", "supporting_evidence": [ { "id": "KF1", "claim": "The correspondence separates distinct problems that are often grouped under multimodal AI.", "context": "Conceptual taxonomy of fusion, missing inputs, time, selective prediction and observation mechanisms.", "evidence": [ { "label": "problem categories", "value": "5" } ], "source_locator": "Corrected proof p. 1, opening paragraph" }, { "id": "KF2", "claim": "The proposed framework makes evidence sufficiency and a defined downstream pathway central to deferral.", "context": "Figure-based conceptual proposal; deferral is distinct from calibration or confidence alone.", "evidence": [ { "label": "principles in figure", "value": "6" }, { "label": "evidence type", "value": "Conceptual framework, not an evaluated intervention" } ], "source_locator": "Corrected proof pp. 2-3; Fig. 1 caption and paragraphs on abstention/deferral" } ] }, { "question": "Which evaluation conditions does it propose?", "answer": "Five candidate conditions are complete inputs, a critical modality missing, natural co-missingness, temporal staleness and geographically or temporally external data.", "evidence_refs": [ "KF3" ], "answer_url": "https://drgezhang.com/papers/doi-10-1016-j-medd-2026-100004.html#qa-3", "source_pointer": "data/deep_geo/doi-10-1016-j-medd-2026-100004.json#/qa/2", "supporting_evidence": [ { "id": "KF3", "claim": "The authors propose intended-use-specific evaluation conditions for incomplete and temporally heterogeneous information.", "context": "Suggested minimum suite requiring future validation and stakeholder consensus.", "evidence": [ { "label": "candidate evaluation conditions", "value": "5: complete inputs, critical modality missing, natural co-missingness, staleness, external data" } ], "source_locator": "Corrected proof p. 3, paragraph beginning Evaluation should be defined by intended use" } ] }, { "question": "Does it report new model accuracy or a clinical trial?", "answer": "No. It is a correspondence presenting conceptual proposals, with no new dataset or validated performance result.", "evidence_refs": [ "KF1", "KF3" ], "answer_url": "https://drgezhang.com/papers/doi-10-1016-j-medd-2026-100004.html#qa-4", "source_pointer": "data/deep_geo/doi-10-1016-j-medd-2026-100004.json#/qa/3", "supporting_evidence": [ { "id": "KF1", "claim": "The correspondence separates distinct problems that are often grouped under multimodal AI.", "context": "Conceptual taxonomy of fusion, missing inputs, time, selective prediction and observation mechanisms.", "evidence": [ { "label": "problem categories", "value": "5" } ], "source_locator": "Corrected proof p. 1, opening paragraph" }, { "id": "KF3", "claim": "The authors propose intended-use-specific evaluation conditions for incomplete and temporally heterogeneous information.", "context": "Suggested minimum suite requiring future validation and stakeholder consensus.", "evidence": [ { "label": "candidate evaluation conditions", "value": "5: complete inputs, critical modality missing, natural co-missingness, staleness, external data" } ], "source_locator": "Corrected proof p. 3, paragraph beginning Evaluation should be defined by intended use" } ] } ], "evidence_scope": { "supports": [ "Attribution of the stated evaluation and deferral proposal to this correspondence.", "Literature discussion of missingness, temporal context and prospective evaluation needs." ], "does_not_establish": [ "Empirical improvement in discrimination, safety, equity or patient outcomes.", "A validated AI algorithm or universal abstention threshold.", "A consensus guideline, clinical standard or independent external validation cohort." ] }, "limitations": [ "Conceptual proposal without empirical validation.", "Evidence locators refer to the archived publisher corrected proof. The reviewed passages and Figure 1 caption match the current issue-assigned publisher HTML after removal of proof page headers and formatting; the current publisher PDF was not independently inspected.", "The five evaluation conditions are candidate recommendations, not consensus criteria." ], "publication": { "authors": [ "Yahui Han", "Qiang Li", "Taiqi Zhao", "Shanshan Cai", "Biyun Li", "Kaisaierjiang Kadier", "Tianshu Gu", "Zhonghua Sun", "Pengyuan Xu", "Dezhi Liu", "Moyan Li", "Xufeng Huang", "Qian Guo", "Yuan Cheng", "Honglin Zheng", "Ge Zhang" ], "year": 2026, "journal": "Medicine Discovery", "type": "Correspondence" } } ], "presentation": "contribution-first-v1" }