Data Analysis Assistant DashboardIntelligent statistics and publication-ready visualization for scientific research

πŸ’‘ Research Hypothesis

A local Scientific Discovery Assistant that asks "what is surprising here?" β€” interactions, discordant biomarkers, unexpected correlations and hidden clusters, ranked by novelty and scientific value, with the statistics behind each finding.

Screened 53 variables Γ— 5 visits Γ— 20 strata in 200 patients (20672 statistical tests) β€” 40 non-obvious findings, 59 obvious/known relationships suppressed.

Minimum novelty

Ranked by scientific value β€” novelty, effect size, statistical robustness, cohort size and confidence β€” not by p-value alone.

#1Interaction paper: VCAM-1 decreases by 6 Month Follow-up only in No AF patients
β˜…β˜…β˜…β˜…β˜…
High confidence
A focused interaction analysis built around the strongest subgroup-restricted effect in this dataset. The manuscript would report the stratified trajectories plus the formal interaction term.
Publication opportunities
n = 342
p = <0.001
effect = -0.94
score 171
Why this is not obvious

Effect exists in one stratum only β€” an interaction, not a known main effect.

Biological plausibility

VCAM-1 reflects innate immune and endothelial activation. A change confined to No AF patients would suggest ongoing immune signalling that is dissociated from acute myocardial necrosis, a mechanism relevant to adverse remodeling.

Statistical evidence
  • No AF: Baseline mean 664.95 β†’ 6 Month Follow-up mean 539.59 ng/mL (n=173/169), p < 0.001
  • AF: no significant change (p = 0.219)
  • Effect size (Cohen's d) = -0.94
  • Stratifying variable: Atrial Fibrillation vs None
  • Candidate figures: Stratified trajectory plot (mean Β± SEM per stratum); Forest plot of stratum-specific effects; Interaction term table
  • Target journals: European Heart Journal – Acute Cardiovascular Care; Clinical Research in Cardiology
Suggested next analyses
  • Test the formal interaction term in a mixed-effects model (variable Γ— subgroup Γ— time)
  • Repeat as multivariable regression adjusted for age, sex and infarct size
  • Pre-specify the subgroup and validate in an external cohort
#2VCAM-1 decreases by 6 Month Follow-up only in No AF patients
β˜…β˜…β˜…β˜…β˜…
High confidence
A significant decrease of VCAM-1 from Baseline to 6 Month Follow-up was present only in No AF patients (p < 0.001), while AF patients showed no change (p = 0.219). The effect is therefore subgroup-restricted rather than a general cohort trend.
Unexpected subgroup & interaction effects
n = 342
p = <0.001
effect = -0.94
score 166
Why this is not obvious

Effect exists in one stratum only β€” an interaction, not a known main effect.

Biological plausibility

VCAM-1 reflects innate immune and endothelial activation. A change confined to No AF patients would suggest ongoing immune signalling that is dissociated from acute myocardial necrosis, a mechanism relevant to adverse remodeling.

Statistical evidence
  • No AF: Baseline mean 664.95 β†’ 6 Month Follow-up mean 539.59 ng/mL (n=173/169), p < 0.001
  • AF: no significant change (p = 0.219)
  • Effect size (Cohen's d) = -0.94
  • Stratifying variable: Atrial Fibrillation vs None
Suggested next analyses
  • Test the formal interaction term in a mixed-effects model (variable Γ— subgroup Γ— time)
  • Repeat as multivariable regression adjusted for age, sex and infarct size
  • Pre-specify the subgroup and validate in an external cohort
#3Heart rate decreases by 6 Month Follow-up only in No AF patients
β˜…β˜…β˜…β˜…β˜…
High confidence
A significant decrease of Heart rate from Baseline to 6 Month Follow-up was present only in No AF patients (p < 0.001), while AF patients showed no change (p = 0.234). The effect is therefore subgroup-restricted rather than a general cohort trend.
Unexpected subgroup & interaction effects
n = 347
p = <0.001
effect = -0.91
score 166
Why this is not obvious

Effect exists in one stratum only β€” an interaction, not a known main effect.

Biological plausibility

A change confined to No AF patients may reflect a phenotype-specific biological pathway worth mechanistic investigation.

Statistical evidence
  • No AF: Baseline mean 84.18 β†’ 6 Month Follow-up mean 72.70 bpm (n=173/174), p < 0.001
  • AF: no significant change (p = 0.234)
  • Effect size (Cohen's d) = -0.91
  • Stratifying variable: Atrial Fibrillation vs None
Suggested next analyses
  • Test the formal interaction term in a mixed-effects model (variable Γ— subgroup Γ— time)
  • Repeat as multivariable regression adjusted for age, sex and infarct size
  • Pre-specify the subgroup and validate in an external cohort
#4CK-MB decreases by Day 5 only in No Anemia patients
β˜…β˜…β˜…β˜…β˜…
High confidence
A significant decrease of CK-MB from Baseline to Day 5 was present only in No Anemia patients (p < 0.001), while Anemia patients showed no change (p = 0.276). The effect is therefore subgroup-restricted rather than a general cohort trend.
Unexpected subgroup & interaction effects
n = 312
p = <0.001
effect = -0.87
score 165
Why this is not obvious

Effect exists in one stratum only β€” an interaction, not a known main effect.

Biological plausibility

CK-MB indexes myocyte injury and wall stress. A change confined to No Anemia patients may reflect ongoing micro-injury or haemodynamic load rather than the index infarct.

Statistical evidence
  • No Anemia: Baseline mean 22.90 β†’ Day 5 mean 11.84 ng/mL (n=156/156), p < 0.001
  • Anemia: no significant change (p = 0.276)
  • Effect size (Cohen's d) = -0.87
  • Stratifying variable: Anemia vs No Anemia
Suggested next analyses
  • Test the formal interaction term in a mixed-effects model (variable Γ— subgroup Γ— time)
  • Repeat as multivariable regression adjusted for age, sex and infarct size
  • Pre-specify the subgroup and validate in an external cohort
#5Cardiac Output normalises while Stroke Volume stays elevated β€” despite being coupled at baseline
β˜…β˜…β˜…β˜…β˜…
High confidence
Cardiac Output and Stroke Volume are correlated at Baseline (r = 0.94, p < 0.001) yet uncouple over follow-up: Cardiac Output falls significantly by 1 Year Follow-up while Stroke Volume does not change.
Discordant biomarkers
n = 180
p = <0.001
effect = 0.94
score 165
Why this is not obvious

Two coupled markers dissociate over time β€” a discordance, not an expected trend.

Biological plausibility

Two markers that share a baseline pathway but separate over time suggest distinct downstream biology β€” typically resolution of acute injury (Cardiac Output) versus a chronic, self-sustaining process (Stroke Volume).

Statistical evidence
  • Baseline coupling: Pearson r = 0.94 (n = 180), p < 0.001
  • Cardiac Output: significant decrease Baseline β†’ 1 Year Follow-up
  • Stroke Volume: no significant change over the same interval
Suggested next analyses
  • Model both trajectories jointly (bivariate mixed-effects model)
  • Test whether the residual (discordance) predicts outcome
  • Perform mediation analysis to test the intermediate pathway
  • Confirm with non-parametric (Spearman) and bootstrap confidence intervals
#6E/e' decreases by 1 Year Follow-up only in No CKD patients
β˜…β˜…β˜…β˜…β˜…
High confidence
A significant decrease of E/e' from Baseline to 1 Year Follow-up was present only in No CKD patients (p < 0.001), while CKD patients showed no change (p = 0.227). The effect is therefore subgroup-restricted rather than a general cohort trend.
Unexpected subgroup & interaction effects
n = 316
p = <0.001
effect = -0.83
score 164
Why this is not obvious

Effect exists in one stratum only β€” an interaction, not a known main effect.

Biological plausibility

E/e' is a structural/functional remodeling measure; A change confined to No CKD patients may indicate a distinct healing phenotype (fibrosis, oedema resolution or microvascular obstruction).

Statistical evidence
  • No CKD: Baseline mean 11.76 β†’ 1 Year Follow-up mean 9.12 (n=161/155), p < 0.001
  • CKD: no significant change (p = 0.227)
  • Effect size (Cohen's d) = -0.83
  • Stratifying variable: CKD vs No CKD
Suggested next analyses
  • Test the formal interaction term in a mixed-effects model (variable Γ— subgroup Γ— time)
  • Repeat as multivariable regression adjusted for age, sex and infarct size
  • Pre-specify the subgroup and validate in an external cohort
#7QRS duration decreases by 6 Month Follow-up only in High BNP patients
β˜…β˜…β˜…β˜…β˜…
High confidence
A significant decrease of QRS duration from Baseline to 6 Month Follow-up was present only in High BNP patients (p < 0.001), while Low BNP patients showed no change (p = 0.488). The effect is therefore subgroup-restricted rather than a general cohort trend.
Unexpected subgroup & interaction effects
n = 94
p = <0.001
effect = -1.11
score 163
Why this is not obvious

Effect exists in one stratum only β€” an interaction, not a known main effect.

Biological plausibility

A change confined to High BNP patients may reflect a phenotype-specific biological pathway worth mechanistic investigation.

Statistical evidence
  • High BNP: Baseline mean 115.10 β†’ 6 Month Follow-up mean 97.80 ms (n=48/46), p < 0.001
  • Low BNP: no significant change (p = 0.488)
  • Effect size (Cohen's d) = -1.11
  • Stratifying variable: High NT-proBNP vs Low NT-proBNP (median 540)
Suggested next analyses
  • Test the formal interaction term in a mixed-effects model (variable Γ— subgroup Γ— time)
  • Repeat as multivariable regression adjusted for age, sex and infarct size
  • Pre-specify the subgroup and validate in an external cohort
#8Effect-modification paper: Troponin correlates with Native T1 only in Non-Anterior patients
β˜…β˜…β˜…β˜…β˜…
High confidence
A biomarker–imaging association that is present in one stratum only supports a short, figure-driven manuscript on effect modification.
Publication opportunities
n = 145
p = <0.001
effect = 0.52
score 162
Why this is not obvious

Cross-domain association present in one stratum only (effect modification).

Biological plausibility

Native T1 is a structural/functional remodeling measure; A biomarker–imaging coupling restricted to Non-Anterior patients may indicate a distinct healing phenotype (fibrosis, oedema resolution or microvascular obstruction).

Statistical evidence
  • Non-Anterior: r = 0.52, n = 145, p < 0.001
  • Anterior: r = -0.05, n = 40, p = 0.765
  • Effect modifier: Anterior MI vs Non-Anterior MI
  • Cross-domain association (Cardiac biomarkers Γ— MRI)
  • Candidate figures: Scatter plot with separate regression lines per stratum; Correlation heat map by stratum
  • Target journals: International Journal of Cardiology; Journal of Clinical Medicine
Suggested next analyses
  • Fit multivariable linear regression with the interaction term
  • Perform mediation analysis to test the intermediate pathway
  • Confirm with non-parametric (Spearman) and bootstrap confidence intervals
#9QRS duration decreases by 6 Month Follow-up only in Anterior patients
β˜…β˜…β˜…β˜…β˜…
High confidence
A significant decrease of QRS duration from Baseline to 6 Month Follow-up was present only in Anterior patients (p < 0.001), while Non-Anterior patients showed no change (p = 0.704). The effect is therefore subgroup-restricted rather than a general cohort trend.
Unexpected subgroup & interaction effects
n = 86
p = <0.001
effect = -1.05
score 159
Why this is not obvious

Effect exists in one stratum only β€” an interaction, not a known main effect.

Biological plausibility

A change confined to Anterior patients may reflect a phenotype-specific biological pathway worth mechanistic investigation.

Statistical evidence
  • Anterior: Baseline mean 112.72 β†’ 6 Month Follow-up mean 95.37 ms (n=43/43), p < 0.001
  • Non-Anterior: no significant change (p = 0.704)
  • Effect size (Cohen's d) = -1.05
  • Stratifying variable: Anterior MI vs Non-Anterior MI
Suggested next analyses
  • Test the formal interaction term in a mixed-effects model (variable Γ— subgroup Γ— time)
  • Repeat as multivariable regression adjusted for age, sex and infarct size
  • Pre-specify the subgroup and validate in an external cohort
#10TAPSE increases by 6 Month Follow-up only in High BNP patients
β˜…β˜…β˜…β˜…β˜…
High confidence
A significant increase of TAPSE from Baseline to 6 Month Follow-up was present only in High BNP patients (p < 0.001), while Low BNP patients showed no change (p = 0.818). The effect is therefore subgroup-restricted rather than a general cohort trend.
Unexpected subgroup & interaction effects
n = 98
p = <0.001
effect = 0.96
score 157
Why this is not obvious

Effect exists in one stratum only β€” an interaction, not a known main effect.

Biological plausibility

TAPSE is a structural/functional remodeling measure; A change confined to High BNP patients may indicate a distinct healing phenotype (fibrosis, oedema resolution or microvascular obstruction).

Statistical evidence
  • High BNP: Baseline mean 17.67 β†’ 6 Month Follow-up mean 20.47 mm (n=49/49), p < 0.001
  • Low BNP: no significant change (p = 0.818)
  • Effect size (Cohen's d) = 0.96
  • Stratifying variable: High NT-proBNP vs Low NT-proBNP (median 540)
Suggested next analyses
  • Test the formal interaction term in a mixed-effects model (variable Γ— subgroup Γ— time)
  • Repeat as multivariable regression adjusted for age, sex and infarct size
  • Pre-specify the subgroup and validate in an external cohort
#11Troponin correlates with Native T1 only in Non-Anterior patients
β˜…β˜…β˜…β˜…β˜…
High confidence
Troponin was associated with Native T1 in Non-Anterior patients (r = 0.52, p < 0.001) but not in Anterior patients (r = -0.05, p = 0.765) β€” an effect-modification pattern rather than a general association.
Unexpected correlations
n = 145
p = <0.001
effect = 0.52
score 157
Why this is not obvious

Cross-domain association present in one stratum only (effect modification).

Biological plausibility

Native T1 is a structural/functional remodeling measure; A biomarker–imaging coupling restricted to Non-Anterior patients may indicate a distinct healing phenotype (fibrosis, oedema resolution or microvascular obstruction).

Statistical evidence
  • Non-Anterior: r = 0.52, n = 145, p < 0.001
  • Anterior: r = -0.05, n = 40, p = 0.765
  • Effect modifier: Anterior MI vs Non-Anterior MI
  • Cross-domain association (Cardiac biomarkers Γ— MRI)
Suggested next analyses
  • Fit multivariable linear regression with the interaction term
  • Perform mediation analysis to test the intermediate pathway
  • Confirm with non-parametric (Spearman) and bootstrap confidence intervals
#12Troponin correlates with T2 mapping only in Non-Anterior patients
β˜…β˜…β˜…β˜…β˜…
High confidence
Troponin was associated with T2 mapping in Non-Anterior patients (r = 0.51, p < 0.001) but not in Anterior patients (r = 0.08, p = 0.639) β€” an effect-modification pattern rather than a general association.
Unexpected correlations
n = 142
p = <0.001
effect = 0.51
score 157
Why this is not obvious

Cross-domain association present in one stratum only (effect modification).

Biological plausibility

T2 mapping is a structural/functional remodeling measure; A biomarker–imaging coupling restricted to Non-Anterior patients may indicate a distinct healing phenotype (fibrosis, oedema resolution or microvascular obstruction).

Statistical evidence
  • Non-Anterior: r = 0.51, n = 142, p < 0.001
  • Anterior: r = 0.08, n = 40, p = 0.639
  • Effect modifier: Anterior MI vs Non-Anterior MI
  • Cross-domain association (Cardiac biomarkers Γ— MRI)
Suggested next analyses
  • Fit multivariable linear regression with the interaction term
  • Perform mediation analysis to test the intermediate pathway
  • Confirm with non-parametric (Spearman) and bootstrap confidence intervals
#13CRP correlates with Native T1 only in Non-Anterior patients
β˜…β˜…β˜…β˜…β˜…
Moderate confidence
CRP was associated with Native T1 in Non-Anterior patients (r = 0.49, p < 0.001) but not in Anterior patients (r = -0.12, p = 0.477) β€” an effect-modification pattern rather than a general association.
Unexpected correlations
n = 144
p = <0.001
effect = 0.49
score 151
Why this is not obvious

Cross-domain association present in one stratum only (effect modification).

Biological plausibility

Native T1 is a structural/functional remodeling measure; A biomarker–imaging coupling restricted to Non-Anterior patients may indicate a distinct healing phenotype (fibrosis, oedema resolution or microvascular obstruction).

Statistical evidence
  • Non-Anterior: r = 0.49, n = 144, p < 0.001
  • Anterior: r = -0.12, n = 40, p = 0.477
  • Effect modifier: Anterior MI vs Non-Anterior MI
  • Cross-domain association (Inflammatory biomarkers Γ— MRI)
Suggested next analyses
  • Fit multivariable linear regression with the interaction term
  • Perform mediation analysis to test the intermediate pathway
  • Confirm with non-parametric (Spearman) and bootstrap confidence intervals
#14CK correlates with Native T1 only in Non-Anterior patients
β˜…β˜…β˜…β˜…β˜…
Moderate confidence
CK was associated with Native T1 in Non-Anterior patients (r = 0.49, p < 0.001) but not in Anterior patients (r = -0.09, p = 0.558) β€” an effect-modification pattern rather than a general association.
Unexpected correlations
n = 144
p = <0.001
effect = 0.49
score 151
Why this is not obvious

Cross-domain association present in one stratum only (effect modification).

Biological plausibility

Native T1 is a structural/functional remodeling measure; A biomarker–imaging coupling restricted to Non-Anterior patients may indicate a distinct healing phenotype (fibrosis, oedema resolution or microvascular obstruction).

Statistical evidence
  • Non-Anterior: r = 0.49, n = 144, p < 0.001
  • Anterior: r = -0.09, n = 43, p = 0.558
  • Effect modifier: Anterior MI vs Non-Anterior MI
  • Cross-domain association (Cardiac biomarkers Γ— MRI)
Suggested next analyses
  • Fit multivariable linear regression with the interaction term
  • Perform mediation analysis to test the intermediate pathway
  • Confirm with non-parametric (Spearman) and bootstrap confidence intervals
#15IL-6 correlates with Native T1 only in Non-Anterior patients
β˜…β˜…β˜…β˜…β˜…
Moderate confidence
IL-6 was associated with Native T1 in Non-Anterior patients (r = 0.49, p < 0.001) but not in Anterior patients (r = -0.14, p = 0.392) β€” an effect-modification pattern rather than a general association.
Unexpected correlations
n = 145
p = <0.001
effect = 0.49
score 151
Why this is not obvious

Cross-domain association present in one stratum only (effect modification).

Biological plausibility

Native T1 is a structural/functional remodeling measure; A biomarker–imaging coupling restricted to Non-Anterior patients may indicate a distinct healing phenotype (fibrosis, oedema resolution or microvascular obstruction).

Statistical evidence
  • Non-Anterior: r = 0.49, n = 145, p < 0.001
  • Anterior: r = -0.14, n = 41, p = 0.392
  • Effect modifier: Anterior MI vs Non-Anterior MI
  • Cross-domain association (Inflammatory biomarkers Γ— MRI)
Suggested next analyses
  • Fit multivariable linear regression with the interaction term
  • Perform mediation analysis to test the intermediate pathway
  • Confirm with non-parametric (Spearman) and bootstrap confidence intervals
#16CK-MB correlates with Native T1 only in Non-Anterior patients
β˜…β˜…β˜…β˜…β˜…
Moderate confidence
CK-MB was associated with Native T1 in Non-Anterior patients (r = 0.48, p < 0.001) but not in Anterior patients (r = -0.14, p = 0.369) β€” an effect-modification pattern rather than a general association.
Unexpected correlations
n = 140
p = <0.001
effect = 0.48
score 150
Why this is not obvious

Cross-domain association present in one stratum only (effect modification).

Biological plausibility

Native T1 is a structural/functional remodeling measure; A biomarker–imaging coupling restricted to Non-Anterior patients may indicate a distinct healing phenotype (fibrosis, oedema resolution or microvascular obstruction).

Statistical evidence
  • Non-Anterior: r = 0.48, n = 140, p < 0.001
  • Anterior: r = -0.14, n = 42, p = 0.369
  • Effect modifier: Anterior MI vs Non-Anterior MI
  • Cross-domain association (Cardiac biomarkers Γ— MRI)
Suggested next analyses
  • Fit multivariable linear regression with the interaction term
  • Perform mediation analysis to test the intermediate pathway
  • Confirm with non-parametric (Spearman) and bootstrap confidence intervals
#17CK-MB correlates with Infarct size only in Non-Anterior patients
β˜…β˜…β˜…β˜…β˜…
Moderate confidence
CK-MB was associated with Infarct size in Non-Anterior patients (r = 0.46, p < 0.001) but not in Anterior patients (r = -0.04, p = 0.818) β€” an effect-modification pattern rather than a general association.
Unexpected correlations
n = 144
p = <0.001
effect = 0.46
score 150
Why this is not obvious

Cross-domain association present in one stratum only (effect modification).

Biological plausibility

Infarct size is a structural/functional remodeling measure; A biomarker–imaging coupling restricted to Non-Anterior patients may indicate a distinct healing phenotype (fibrosis, oedema resolution or microvascular obstruction).

Statistical evidence
  • Non-Anterior: r = 0.46, n = 144, p < 0.001
  • Anterior: r = -0.04, n = 43, p = 0.818
  • Effect modifier: Anterior MI vs Non-Anterior MI
  • Cross-domain association (Cardiac biomarkers Γ— MRI)
Suggested next analyses
  • Fit multivariable linear regression with the interaction term
  • Perform mediation analysis to test the intermediate pathway
  • Confirm with non-parametric (Spearman) and bootstrap confidence intervals
#18IL-8 correlates with T2 mapping only in Non-Anterior patients
β˜…β˜…β˜…β˜…β˜…
Moderate confidence
IL-8 was associated with T2 mapping in Non-Anterior patients (r = 0.46, p < 0.001) but not in Anterior patients (r = 0.07, p = 0.669) β€” an effect-modification pattern rather than a general association.
Unexpected correlations
n = 142
p = <0.001
effect = 0.46
score 150
Why this is not obvious

Cross-domain association present in one stratum only (effect modification).

Biological plausibility

T2 mapping is a structural/functional remodeling measure; A biomarker–imaging coupling restricted to Non-Anterior patients may indicate a distinct healing phenotype (fibrosis, oedema resolution or microvascular obstruction).

Statistical evidence
  • Non-Anterior: r = 0.46, n = 142, p < 0.001
  • Anterior: r = 0.07, n = 39, p = 0.669
  • Effect modifier: Anterior MI vs Non-Anterior MI
  • Cross-domain association (Inflammatory biomarkers Γ— MRI)
Suggested next analyses
  • Fit multivariable linear regression with the interaction term
  • Perform mediation analysis to test the intermediate pathway
  • Confirm with non-parametric (Spearman) and bootstrap confidence intervals
#19CK-MB correlates with MVO extent only in Non-Anterior patients
β˜…β˜…β˜…β˜…β˜…
Moderate confidence
CK-MB was associated with MVO extent in Non-Anterior patients (r = 0.46, p < 0.001) but not in Anterior patients (r = -0.11, p = 0.481) β€” an effect-modification pattern rather than a general association.
Unexpected correlations
n = 138
p = <0.001
effect = 0.46
score 150
Why this is not obvious

Cross-domain association present in one stratum only (effect modification).

Biological plausibility

MVO extent is a structural/functional remodeling measure; A biomarker–imaging coupling restricted to Non-Anterior patients may indicate a distinct healing phenotype (fibrosis, oedema resolution or microvascular obstruction).

Statistical evidence
  • Non-Anterior: r = 0.46, n = 138, p < 0.001
  • Anterior: r = -0.11, n = 44, p = 0.481
  • Effect modifier: Anterior MI vs Non-Anterior MI
  • Cross-domain association (Cardiac biomarkers Γ— MRI)
Suggested next analyses
  • Fit multivariable linear regression with the interaction term
  • Perform mediation analysis to test the intermediate pathway
  • Confirm with non-parametric (Spearman) and bootstrap confidence intervals
#20IL-8 correlates with Native T1 only in Non-Anterior patients
β˜…β˜…β˜…β˜…β˜…
Moderate confidence
IL-8 was associated with Native T1 in Non-Anterior patients (r = 0.43, p < 0.001) but not in Anterior patients (r = 0.08, p = 0.612) β€” an effect-modification pattern rather than a general association.
Unexpected correlations
n = 146
p = <0.001
effect = 0.43
score 150
Why this is not obvious

Cross-domain association present in one stratum only (effect modification).

Biological plausibility

Native T1 is a structural/functional remodeling measure; A biomarker–imaging coupling restricted to Non-Anterior patients may indicate a distinct healing phenotype (fibrosis, oedema resolution or microvascular obstruction).

Statistical evidence
  • Non-Anterior: r = 0.43, n = 146, p < 0.001
  • Anterior: r = 0.08, n = 39, p = 0.612
  • Effect modifier: Anterior MI vs Non-Anterior MI
  • Cross-domain association (Inflammatory biomarkers Γ— MRI)
Suggested next analyses
  • Fit multivariable linear regression with the interaction term
  • Perform mediation analysis to test the intermediate pathway
  • Confirm with non-parametric (Spearman) and bootstrap confidence intervals
#21CK-MB correlates with T2 mapping only in Non-Anterior patients
β˜…β˜…β˜…β˜…β˜…
Moderate confidence
CK-MB was associated with T2 mapping in Non-Anterior patients (r = 0.46, p < 0.001) but not in Anterior patients (r = 0.15, p = 0.340) β€” an effect-modification pattern rather than a general association.
Unexpected correlations
n = 137
p = <0.001
effect = 0.46
score 150
Why this is not obvious

Cross-domain association present in one stratum only (effect modification).

Biological plausibility

T2 mapping is a structural/functional remodeling measure; A biomarker–imaging coupling restricted to Non-Anterior patients may indicate a distinct healing phenotype (fibrosis, oedema resolution or microvascular obstruction).

Statistical evidence
  • Non-Anterior: r = 0.46, n = 137, p < 0.001
  • Anterior: r = 0.15, n = 42, p = 0.340
  • Effect modifier: Anterior MI vs Non-Anterior MI
  • Cross-domain association (Cardiac biomarkers Γ— MRI)
Suggested next analyses
  • Fit multivariable linear regression with the interaction term
  • Perform mediation analysis to test the intermediate pathway
  • Confirm with non-parametric (Spearman) and bootstrap confidence intervals
#22CK correlates with T2 mapping only in Non-Anterior patients
β˜…β˜…β˜…β˜…β˜…
Moderate confidence
CK was associated with T2 mapping in Non-Anterior patients (r = 0.44, p < 0.001) but not in Anterior patients (r = 0.14, p = 0.375) β€” an effect-modification pattern rather than a general association.
Unexpected correlations
n = 142
p = <0.001
effect = 0.44
score 150
Why this is not obvious

Cross-domain association present in one stratum only (effect modification).

Biological plausibility

T2 mapping is a structural/functional remodeling measure; A biomarker–imaging coupling restricted to Non-Anterior patients may indicate a distinct healing phenotype (fibrosis, oedema resolution or microvascular obstruction).

Statistical evidence
  • Non-Anterior: r = 0.44, n = 142, p < 0.001
  • Anterior: r = 0.14, n = 43, p = 0.375
  • Effect modifier: Anterior MI vs Non-Anterior MI
  • Cross-domain association (Cardiac biomarkers Γ— MRI)
Suggested next analyses
  • Fit multivariable linear regression with the interaction term
  • Perform mediation analysis to test the intermediate pathway
  • Confirm with non-parametric (Spearman) and bootstrap confidence intervals
#23Prognostic trajectory paper
β˜…β˜…β˜…β˜…β˜†
High confidence
A manuscript centred on markers that separate outcome groups only during follow-up, arguing for serial rather than single-timepoint sampling.
Publication opportunities
n = 176
p = 0.002
effect = 0.65
score 149
Why this is not obvious

Prognostic separation emerges only during follow-up, not at presentation.

Biological plausibility

ECV is a structural/functional remodeling measure; Late divergence rather than an acute-phase difference may indicate a distinct healing phenotype (fibrosis, oedema resolution or microvascular obstruction).

Statistical evidence
  • Event group at 1 Year Follow-up: mean 28.75 % (n=36)
  • Event-free group: mean 26.34 % (n=140)
  • Welch t-test p = 0.002; Cohen's d = 0.65
  • No baseline difference (p = 0.076) β€” the signal is temporal, not baseline risk
  • Candidate figures: Trajectory plot by outcome group; Time-dependent ROC curves; Kaplan-Meier by follow-up tertile
  • Target journals: American Journal of Cardiology; Open Heart
Suggested next analyses
  • Build a ROC / AUC model and identify an optimal threshold
  • Fit a Cox model with time-dependent covariates
  • Validate in an independent cohort before any clinical claim
#24ECV stays elevated at 1 Year Follow-up only in patients with a later event
β˜…β˜…β˜…β˜…β˜…
High confidence
ECV did not differ between event and event-free patients at Baseline (p = 0.076), but diverged by 1 Year Follow-up (p = 0.002). The separation appears during follow-up rather than at presentation.
Unexpected longitudinal patterns
n = 176
p = 0.002
effect = 0.65
score 144
Why this is not obvious

Prognostic separation emerges only during follow-up, not at presentation.

Biological plausibility

ECV is a structural/functional remodeling measure; Late divergence rather than an acute-phase difference may indicate a distinct healing phenotype (fibrosis, oedema resolution or microvascular obstruction).

Statistical evidence
  • Event group at 1 Year Follow-up: mean 28.75 % (n=36)
  • Event-free group: mean 26.34 % (n=140)
  • Welch t-test p = 0.002; Cohen's d = 0.65
  • No baseline difference (p = 0.076) β€” the signal is temporal, not baseline risk
Suggested next analyses
  • Build a ROC / AUC model and identify an optimal threshold
  • Fit a Cox model with time-dependent covariates
  • Validate in an independent cohort before any clinical claim
#25LVESD decreases by 6 Month Follow-up only in No Previous MI patients
β˜…β˜…β˜…β˜…β˜†
High confidence
A significant decrease of LVESD from Baseline to 6 Month Follow-up was present only in No Previous MI patients (p < 0.001), while Previous MI patients showed no change (p = 0.209). The effect is therefore subgroup-restricted rather than a general cohort trend.
Unexpected subgroup & interaction effects
n = 319
p = <0.001
effect = -0.71
score 143
Why this is not obvious

Effect exists in one stratum only β€” an interaction, not a known main effect.

Biological plausibility

LVESD is a structural/functional remodeling measure; A change confined to No Previous MI patients may indicate a distinct healing phenotype (fibrosis, oedema resolution or microvascular obstruction).

Statistical evidence
  • No Previous MI: Baseline mean 39.56 β†’ 6 Month Follow-up mean 35.75 mm (n=161/158), p < 0.001
  • Previous MI: no significant change (p = 0.209)
  • Effect size (Cohen's d) = -0.71
  • Stratifying variable: Previous MI vs No Previous MI
Suggested next analyses
  • Test the formal interaction term in a mixed-effects model (variable Γ— subgroup Γ— time)
  • Repeat as multivariable regression adjusted for age, sex and infarct size
  • Pre-specify the subgroup and validate in an external cohort
#26LVESD decreases by 6 Month Follow-up only in No AF patients
β˜…β˜…β˜…β˜…β˜†
High confidence
A significant decrease of LVESD from Baseline to 6 Month Follow-up was present only in No AF patients (p < 0.001), while AF patients showed no change (p = 0.277). The effect is therefore subgroup-restricted rather than a general cohort trend.
Unexpected subgroup & interaction effects
n = 338
p = <0.001
effect = -0.69
score 142
Why this is not obvious

Effect exists in one stratum only β€” an interaction, not a known main effect.

Biological plausibility

LVESD is a structural/functional remodeling measure; A change confined to No AF patients may indicate a distinct healing phenotype (fibrosis, oedema resolution or microvascular obstruction).

Statistical evidence
  • No AF: Baseline mean 39.56 β†’ 6 Month Follow-up mean 35.90 mm (n=173/165), p < 0.001
  • AF: no significant change (p = 0.277)
  • Effect size (Cohen's d) = -0.69
  • Stratifying variable: Atrial Fibrillation vs None
Suggested next analyses
  • Test the formal interaction term in a mixed-effects model (variable Γ— subgroup Γ— time)
  • Repeat as multivariable regression adjusted for age, sex and infarct size
  • Pre-specify the subgroup and validate in an external cohort
#27Leukocytes increases by Day 1 only in No HF patients
β˜…β˜…β˜…β˜…β˜†
High confidence
A significant increase of Leukocytes from Baseline to Day 1 was present only in No HF patients (p < 0.001), while HF patients showed no change (p = 0.886). The effect is therefore subgroup-restricted rather than a general cohort trend.
Unexpected subgroup & interaction effects
n = 332
p = <0.001
effect = 0.68
score 142
Why this is not obvious

Effect exists in one stratum only β€” an interaction, not a known main effect.

Biological plausibility

A change confined to No HF patients may reflect a phenotype-specific biological pathway worth mechanistic investigation.

Statistical evidence
  • No HF: Baseline mean 11.64 β†’ Day 1 mean 13.27 10Β³/Β΅L (n=165/167), p < 0.001
  • HF: no significant change (p = 0.886)
  • Effect size (Cohen's d) = 0.68
  • Stratifying variable: Heart Failure vs None
Suggested next analyses
  • Test the formal interaction term in a mixed-effects model (variable Γ— subgroup Γ— time)
  • Repeat as multivariable regression adjusted for age, sex and infarct size
  • Pre-specify the subgroup and validate in an external cohort
#28MCP-1 decreases by Day 5 only in Low BNP patients
β˜…β˜…β˜…β˜…β˜†
High confidence
A significant decrease of MCP-1 from Baseline to Day 5 was present only in Low BNP patients (p < 0.001), while High BNP patients showed no change (p = 0.598). The effect is therefore subgroup-restricted rather than a general cohort trend.
Unexpected subgroup & interaction effects
n = 287
p = <0.001
effect = -0.67
score 142
Why this is not obvious

Effect exists in one stratum only β€” an interaction, not a known main effect.

Biological plausibility

MCP-1 reflects innate immune and endothelial activation. A change confined to Low BNP patients would suggest ongoing immune signalling that is dissociated from acute myocardial necrosis, a mechanism relevant to adverse remodeling.

Statistical evidence
  • Low BNP: Baseline mean 367.32 β†’ Day 5 mean 306.02 pg/mL (n=145/142), p < 0.001
  • High BNP: no significant change (p = 0.598)
  • Effect size (Cohen's d) = -0.67
  • Stratifying variable: High NT-proBNP vs Low NT-proBNP (median 540)
Suggested next analyses
  • Test the formal interaction term in a mixed-effects model (variable Γ— subgroup Γ— time)
  • Repeat as multivariable regression adjusted for age, sex and infarct size
  • Pre-specify the subgroup and validate in an external cohort
#29IL-1Ξ² decreases by Day 5 only in No CKD patients
β˜…β˜…β˜…β˜…β˜†
High confidence
A significant decrease of IL-1Ξ² from Baseline to Day 5 was present only in No CKD patients (p < 0.001), while CKD patients showed no change (p = 0.903). The effect is therefore subgroup-restricted rather than a general cohort trend.
Unexpected subgroup & interaction effects
n = 323
p = <0.001
effect = -0.66
score 142
Why this is not obvious

Effect exists in one stratum only β€” an interaction, not a known main effect.

Biological plausibility

IL-1Ξ² reflects innate immune and endothelial activation. A change confined to No CKD patients would suggest ongoing immune signalling that is dissociated from acute myocardial necrosis, a mechanism relevant to adverse remodeling.

Statistical evidence
  • No CKD: Baseline mean 5.95 β†’ Day 5 mean 4.67 pg/mL (n=161/162), p < 0.001
  • CKD: no significant change (p = 0.903)
  • Effect size (Cohen's d) = -0.66
  • Stratifying variable: CKD vs No CKD
Suggested next analyses
  • Test the formal interaction term in a mixed-effects model (variable Γ— subgroup Γ— time)
  • Repeat as multivariable regression adjusted for age, sex and infarct size
  • Pre-specify the subgroup and validate in an external cohort
#30Combined biomarker score / prediction model
β˜…β˜…β˜…β˜…β˜†
Moderate confidence
The top-ranked candidate markers can be combined into a cross-validated score for late ventricular function, reported with discrimination and calibration.
Publication opportunities
n = 187
p = <0.001
effect = -0.32
score 131
Why this is not obvious

Marker outperforms the conventional reference marker for remodeling.

Biological plausibility

IL-8 reflects innate immune and endothelial activation. An early marker tracking with late ventricular function would suggest ongoing immune signalling that is dissociated from acute myocardial necrosis, a mechanism relevant to adverse remodeling.

Statistical evidence
  • Pearson r = -0.32 (n = 187), p < 0.001
  • Rank 2 of 30 screened baseline variables
  • Reference marker CRP: r = -0.25, p < 0.001
  • Candidate figures: ROC curves (single markers vs combined score); Calibration plot; Variable-importance plot
  • Target journals: Journal of the American Heart Association; Biomarkers in Medicine
Suggested next analyses
  • Multivariable model adjusted for infarct size, age and sex
  • Combined biomarker score with cross-validated performance
  • Build a ROC / AUC model and identify an optimal threshold
  • Fit a Cox model with time-dependent covariates
  • Validate in an independent cohort before any clinical claim
#31Echo LVEF stays elevated at 1 Year Follow-up only in patients with a later event
β˜…β˜…β˜…β˜…β˜…
Moderate confidence
Echo LVEF did not differ between event and event-free patients at Baseline (p = 0.094), but diverged by 1 Year Follow-up (p = 0.019). The separation appears during follow-up rather than at presentation.
Unexpected longitudinal patterns
n = 181
p = 0.019
effect = -0.48
score 131
Why this is not obvious

Prognostic separation emerges only during follow-up, not at presentation.

Biological plausibility

Echo LVEF is a structural/functional remodeling measure; Late divergence rather than an acute-phase difference may indicate a distinct healing phenotype (fibrosis, oedema resolution or microvascular obstruction).

Statistical evidence
  • Event group at 1 Year Follow-up: mean 54.98 % (n=39)
  • Event-free group: mean 58.98 % (n=142)
  • Welch t-test p = 0.019; Cohen's d = -0.48
  • No baseline difference (p = 0.094) β€” the signal is temporal, not baseline risk
Suggested next analyses
  • Build a ROC / AUC model and identify an optimal threshold
  • Fit a Cox model with time-dependent covariates
  • Validate in an independent cohort before any clinical claim
#32Leukocytes stays elevated at 1 Year Follow-up only in patients with a later event
β˜…β˜…β˜…β˜…β˜…
Moderate confidence
Leukocytes did not differ between event and event-free patients at Baseline (p = 0.697), but diverged by 1 Year Follow-up (p = 0.024). The separation appears during follow-up rather than at presentation.
Unexpected longitudinal patterns
n = 177
p = 0.024
effect = 0.43
score 129
Why this is not obvious

Prognostic separation emerges only during follow-up, not at presentation.

Biological plausibility

Late divergence rather than an acute-phase difference may reflect a phenotype-specific biological pathway worth mechanistic investigation.

Statistical evidence
  • Event group at 1 Year Follow-up: mean 7.97 10Β³/Β΅L (n=35)
  • Event-free group: mean 7.02 10Β³/Β΅L (n=142)
  • Welch t-test p = 0.024; Cohen's d = 0.43
  • No baseline difference (p = 0.697) β€” the signal is temporal, not baseline risk
Suggested next analyses
  • Build a ROC / AUC model and identify an optimal threshold
  • Fit a Cox model with time-dependent covariates
  • Validate in an independent cohort before any clinical claim
#33LA volume stays elevated at 1 Year Follow-up only in patients with a later event
β˜…β˜…β˜…β˜…β˜…
Moderate confidence
LA volume did not differ between event and event-free patients at Baseline (p = 0.865), but diverged by 1 Year Follow-up (p = 0.049). The separation appears during follow-up rather than at presentation.
Unexpected longitudinal patterns
n = 178
p = 0.049
effect = 0.38
score 127
Why this is not obvious

Prognostic separation emerges only during follow-up, not at presentation.

Biological plausibility

LA volume is a structural/functional remodeling measure; Late divergence rather than an acute-phase difference may indicate a distinct healing phenotype (fibrosis, oedema resolution or microvascular obstruction).

Statistical evidence
  • Event group at 1 Year Follow-up: mean 56.41 mL (n=37)
  • Event-free group: mean 52.13 mL (n=141)
  • Welch t-test p = 0.049; Cohen's d = 0.38
  • No baseline difference (p = 0.865) β€” the signal is temporal, not baseline risk
Suggested next analyses
  • Build a ROC / AUC model and identify an optimal threshold
  • Fit a Cox model with time-dependent covariates
  • Validate in an independent cohort before any clinical claim
#34IL-8 ranks among the strongest baseline predictors of late MRI LVEF
β˜…β˜…β˜…β˜…β˜†
Moderate confidence
Baseline IL-8 was associated with MRI LVEF at 1 Year Follow-up (r = -0.32, p < 0.001), a stronger association than CRP (r = -0.25).
Candidate biomarkers
n = 187
p = <0.001
effect = -0.32
score 126
Why this is not obvious

Marker outperforms the conventional reference marker for remodeling.

Biological plausibility

IL-8 reflects innate immune and endothelial activation. An early marker tracking with late ventricular function would suggest ongoing immune signalling that is dissociated from acute myocardial necrosis, a mechanism relevant to adverse remodeling.

Statistical evidence
  • Pearson r = -0.32 (n = 187), p < 0.001
  • Rank 2 of 30 screened baseline variables
  • Reference marker CRP: r = -0.25, p < 0.001
Suggested next analyses
  • Multivariable model adjusted for infarct size, age and sex
  • Combined biomarker score with cross-validated performance
  • Build a ROC / AUC model and identify an optimal threshold
  • Fit a Cox model with time-dependent covariates
  • Validate in an independent cohort before any clinical claim
#35CD40 ranks among the strongest baseline predictors of late MRI LVEF
β˜…β˜…β˜…β˜…β˜†
Moderate confidence
Baseline CD40 was associated with MRI LVEF at 1 Year Follow-up (r = -0.32, p < 0.001), a stronger association than CRP (r = -0.25).
Candidate biomarkers
n = 190
p = <0.001
effect = -0.32
score 126
Why this is not obvious

Marker outperforms the conventional reference marker for remodeling.

Biological plausibility

CD40 reflects innate immune and endothelial activation. An early marker tracking with late ventricular function would suggest ongoing immune signalling that is dissociated from acute myocardial necrosis, a mechanism relevant to adverse remodeling.

Statistical evidence
  • Pearson r = -0.32 (n = 190), p < 0.001
  • Rank 3 of 30 screened baseline variables
  • Reference marker CRP: r = -0.25, p < 0.001
Suggested next analyses
  • Multivariable model adjusted for infarct size, age and sex
  • Combined biomarker score with cross-validated performance
  • Build a ROC / AUC model and identify an optimal threshold
  • Fit a Cox model with time-dependent covariates
  • Validate in an independent cohort before any clinical claim
#36Troponin ranks among the strongest baseline predictors of late MRI LVEF
β˜…β˜…β˜…β˜…β˜†
Moderate confidence
Baseline Troponin was associated with MRI LVEF at 1 Year Follow-up (r = -0.25, p < 0.001), a stronger association than CRP (r = -0.25).
Candidate biomarkers
n = 188
p = <0.001
effect = -0.25
score 116
Why this is not obvious

Marker outperforms the conventional reference marker for remodeling.

Biological plausibility

Troponin indexes myocyte injury and wall stress. An early marker tracking with late ventricular function may reflect ongoing micro-injury or haemodynamic load rather than the index infarct.

Statistical evidence
  • Pearson r = -0.25 (n = 188), p < 0.001
  • Rank 4 of 30 screened baseline variables
  • Reference marker CRP: r = -0.25, p < 0.001
Suggested next analyses
  • Multivariable model adjusted for infarct size, age and sex
  • Combined biomarker score with cross-validated performance
  • Build a ROC / AUC model and identify an optimal threshold
  • Fit a Cox model with time-dependent covariates
  • Validate in an independent cohort before any clinical claim
#37Phenotype paper: unsupervised patient clusters
β˜…β˜…β˜…β˜…β˜…
Moderate confidence
A data-driven phenotyping manuscript describing clusters derived from baseline laboratory and imaging profiles and their outcome differences.
Publication opportunities
n = 32
effect = -0.09
score 114
Why this is not obvious

Data-driven phenotype not defined by any pre-specified clinical variable.

Biological plausibility

A profile combining opposite-direction markers does not follow a single severity axis and may represent a distinct biological phenotype (for example inflammation-dominant versus necrosis-dominant healing) rather than simply 'sicker' patients.

Statistical evidence
  • Cluster size: 32 / 149 patients with complete baseline data
  • Native T1: cluster mean z = -0.80
  • LA volume: cluster mean z = 0.78
  • CD40: cluster mean z = -0.75
  • Infarct size: cluster mean z = -0.72
  • CK: cluster mean z = -0.64
  • Event rate 13% vs 21% (rest of cohort)
  • Candidate figures: Cluster heat map of standardised variables; PCA/UMAP-style cluster projection; Kaplan-Meier by cluster
  • Target journals: Frontiers in Cardiovascular Medicine; Scientific Reports
Suggested next analyses
  • Confirm cluster stability (hierarchical clustering, silhouette, bootstrap)
  • Compare outcomes across clusters (Kaplan-Meier, log-rank)
  • Characterise clusters in multivariable models and test a combined biomarker score
#38Unsupervised cluster 3: ↓ Native T1 / ↑ LA volume / ↓ CD40
β˜…β˜…β˜…β˜…β˜…
Moderate confidence
k-means clustering of 10 standardised baseline variables isolated 32 patients characterised by low Native T1 (z = -0.80), high LA volume (z = 0.78), low CD40 (z = -0.75), low Infarct size (z = -0.72), low CK (z = -0.64), with an event rate of 13% versus 21% in the remaining cohort.
Hidden patient clusters
n = 32
effect = -0.09
score 109
Why this is not obvious

Data-driven phenotype not defined by any pre-specified clinical variable.

Biological plausibility

A profile combining opposite-direction markers does not follow a single severity axis and may represent a distinct biological phenotype (for example inflammation-dominant versus necrosis-dominant healing) rather than simply 'sicker' patients.

Statistical evidence
  • Cluster size: 32 / 149 patients with complete baseline data
  • Native T1: cluster mean z = -0.80
  • LA volume: cluster mean z = 0.78
  • CD40: cluster mean z = -0.75
  • Infarct size: cluster mean z = -0.72
  • CK: cluster mean z = -0.64
  • Event rate 13% vs 21% (rest of cohort)
Suggested next analyses
  • Confirm cluster stability (hierarchical clustering, silhouette, bootstrap)
  • Compare outcomes across clusters (Kaplan-Meier, log-rank)
  • Characterise clusters in multivariable models and test a combined biomarker score
#39Unsupervised cluster 1: ↑ CK / ↑ Native T1 / ↑ Infarct size
β˜…β˜…β˜…β˜…β˜†
Moderate confidence
k-means clustering of 10 standardised baseline variables isolated 77 patients characterised by high CK (z = 0.71), high Native T1 (z = 0.69), high Infarct size (z = 0.68), high CD40 (z = 0.58), with an event rate of 26% versus 13% in the remaining cohort.
Hidden patient clusters
n = 77
effect = 0.13
score 92
Why this is not obvious

Data-driven phenotype not defined by any pre-specified clinical variable.

Biological plausibility

This cluster may capture a coherent severity or inflammatory phenotype that conventional single-variable cut-offs do not identify.

Statistical evidence
  • Cluster size: 77 / 149 patients with complete baseline data
  • CK: cluster mean z = 0.71
  • Native T1: cluster mean z = 0.69
  • Infarct size: cluster mean z = 0.68
  • CD40: cluster mean z = 0.58
  • Event rate 26% vs 13% (rest of cohort)
Suggested next analyses
  • Confirm cluster stability (hierarchical clustering, silhouette, bootstrap)
  • Compare outcomes across clusters (Kaplan-Meier, log-rank)
  • Characterise clusters in multivariable models and test a combined biomarker score
#40Unsupervised cluster 2: ↓ LA volume / ↓ Infarct size / ↓ MVO extent
β˜…β˜…β˜…β˜…β˜†
Moderate confidence
k-means clustering of 10 standardised baseline variables isolated 40 patients characterised by low LA volume (z = -0.94), low Infarct size (z = -0.70), low MVO extent (z = -0.68), low CK (z = -0.67), low Native T1 (z = -0.66), with an event rate of 13% versus 22% in the remaining cohort.
Hidden patient clusters
n = 40
effect = -0.10
score 89
Why this is not obvious

Data-driven phenotype not defined by any pre-specified clinical variable.

Biological plausibility

This cluster may capture a coherent severity or inflammatory phenotype that conventional single-variable cut-offs do not identify.

Statistical evidence
  • Cluster size: 40 / 149 patients with complete baseline data
  • LA volume: cluster mean z = -0.94
  • Infarct size: cluster mean z = -0.70
  • MVO extent: cluster mean z = -0.68
  • CK: cluster mean z = -0.67
  • Native T1: cluster mean z = -0.66
  • Event rate 13% vs 22% (rest of cohort)
Suggested next analyses
  • Confirm cluster stability (hierarchical clustering, silhouette, bootstrap)
  • Compare outcomes across clusters (Kaplan-Meier, log-rank)
  • Characterise clusters in multivariable models and test a combined biomarker score