DeepECG-SSL-CLSA-CVD / model_params.json
robertavrammd's picture
Add strict five-fold CVD inference and full-precision risk equations
12be54e verified
Raw History Blame Contribute Delete
2.52 kB
{
"description": "DeepECG-SSL fine-tuned in CLSA for incident CVD by Follow-up 2 (median 5.8 y). Ensemble of 5 cross-validation fold models.",
"input": {
"leads": 12,
"sampling_rate_hz": 250,
"duration_s": 10,
"samples": 2500,
"units": "mV",
"normalisation": "none (amplitude-preserving)",
"lead_order": "I, II, III, aVR, aVL, aVF, V1-V6"
},
"per_fold_platt": {
"0": {
"a": 0.9256001662643487,
"b": -1.2964684731565865
},
"1": {
"a": 0.7505578577075713,
"b": -1.1098622868479822
},
"2": {
"a": 0.6849553209687097,
"b": -1.3305760314820416
},
"3": {
"a": 0.8646998199738609,
"b": -0.9854216020249019
},
"4": {
"a": 0.7763569971752428,
"b": -1.204451555024041
}
},
"model_A_plus": {
"formula": "logit(risk) = intercept + b_ecg*ecg_logit + b_age*age + b_female*female + b_htn*hypertension",
"intercept": -3.8386085094487012,
"b_ecg": 0.6502693992781706,
"b_age": 0.04871459335113484,
"b_female": -0.3648226286135894,
"b_htn": 0.34160917240450983,
"ecg_logit": "logit of the mean over folds of Platt-calibrated fold probabilities",
"age": "years",
"female": "1 if female",
"hypertension": "1 if self-reported physician-diagnosed hypertension"
},
"model_ecg_age_sex_external": {
"formula": "logit(risk) = intercept + b_ecg*ecg_logit_uncal + b_age*age + b_female*female",
"intercept": -4.685164180420436,
"b_ecg": 0.5186974036163096,
"b_age": 0.052019636929571025,
"b_female": -0.3711663016453705,
"ecg_logit_uncal": "logit of the mean over the 5 folds of UNCALIBRATED fold sigmoid outputs",
"use": "validated at MHI (no hypertension variable); fitted on CLSA development out-of-fold predictions, applied unchanged. Overpredicts at MHI (slope 0.81, CITL -0.66): recalibrate locally."
},
"age_sex_reference": {
"intercept": -6.624631360106959,
"b_age": 0.07239711085716635,
"b_female": -0.4616633628268495
},
"performance": {
"clsa_test_Aplus_auroc": "0.733 (0.707-0.759), n=3,906, 394 events",
"clsa_test_FRS_auroc": 0.689,
"clsa_test_PREVENT_auroc": 0.715,
"mhi_ecg_age_sex_auroc": "0.761 (0.751-0.771), n=21,647, 2,190 events; age+sex 0.728"
},
"parameter_provenance": "Full-precision development-only logistic fits reproduced from the original recipe; every coefficient rounds to the original four-decimal coefficient; A+ verified against all 3906 saved test outputs."
}