MinutiumMinutium homeClinician console

Results

What the model actually does.

Every figure on this page is plotted from the evaluation artifacts, on a held-out fold of patients the model never saw during training or calibration. Nothing here is illustrative.

ROC AUC
0.9670
Sensitivity
0.871
Specificity
0.922
Calibration error
0.0188

1,462 held-out records · 912 NORM · 550 MI-associated · minutium-ecg-res1d-0.1.0

Where the two classes land
01445750.000.250.500.751.00threshold 0.43Calibrated probability of MI-associated morphology
True NORMTrue MI-associated

Median probability: NORM 0.021 · MI 0.988

Distribution of calibrated probabilities on the held-out test split, separated by true class. Overlap between the two distributions is the error the headline AUC averages over.

ROC curve
0.00.51.00.00.51.0False positive rateTrue positive rate

AUC 0.9670 · baseline 0.9293

Held-out test split, against the interpretable baseline.

Calibration
0.00.51.00.00.51.0PredictedObserved

ECE 0.0188

Temperature scaling was fitted on the validation fold. The dashed line is the same model before scaling.

Confusion matrix

True negative

841

False positive

71

False negative

71

True positive

479

At threshold 0.430

Operating point chosen on validation to favour sensitivity.

Training
0.940.960.981815bestEpoch
Validation ROC AUC (left)Training loss (relative)

Early stopped at epoch 15; weights taken from epoch 10, the best validation ROC AUC. Seed 1729.

Model versus baseline
MetricNeuralBaseline
ROC AUC0.96700.9293
PR AUC0.95750.9094
Sensitivity0.87090.8945
Specificity0.92210.7829
PPV0.87090.7130
F10.87090.7935
Brier0.06540.0969
ECE0.01880.0298

Both trained on the same patient-level splits.

Worked example

One patient, two and a half years apart.

Both recordings below come from the same de-identified PTB-XL patient in the held-out split, the normal one first. This is the comparison the product is built around.

NORM

Stored baseline recording

PTB-XL record 6902 · 58 · male

Probability

29.1%

29% to 64%

II
Heart rate
95
Signal
excellent
Uncertainty
wide

Leads the model relied on most

IIIIIaVR
Prerecorded retrospective research ECG from a public dataset. This signal did not come from the person shown in the demo. Both recordings come from the same de-identified PTB-XL patient, 911 days (2.5 years) apart, with the normal recording first. Selected from the held-out test split by a fixed rule: among within-patient normal-to-MI pairs that the model classifies correctly and that grade 'good' or better on waveform quality, the pair with the longest interval between recordings.

MI-associated

Later recording, MI-associated

PTB-XL record 11899 · 61 · male

Probability

86.1%

73% to 90%

II
Heart rate
140
Signal
excellent
Uncertainty
moderate

Leads the model relied on most

IIIIIaVR
Prerecorded retrospective research ECG from a public dataset. This signal did not come from the person shown in the demo. Both recordings come from the same de-identified PTB-XL patient, 911 days (2.5 years) apart, with the normal recording first. Selected from the held-out test split by a fixed rule: among within-patient normal-to-MI pairs that the model classifies correctly and that grade 'good' or better on waveform quality, the pair with the longest interval between recordings.

The unflattering results

What the headline number does not tell you.

These four measurements exist because a single AUC on a curated retrospective dataset is the easiest number in medicine to over-read. They are on this page for the same reason.

Positive predictive value against prevalence
0%50%100%2%5%10%20%38%19%37%55%74%87%True prevalence in the monitored populationPPV

Sensitivity and specificity do not move with prevalence. PPV does, and a person at home with chest pain is a far lower-prevalence population than this cohort.

Fewer leads, as a patch would have
  • full 12 lead0.967
  • limb only0.842
  • patch 6 lead0.932
  • patch 3 lead0.810
  • single lead II0.291

An optimistic stand-in for a chest patch. Removing leads keeps the geometry of the remaining ones intact, whereas a real patch changes electrode positions and therefore the lead vectors themselves.

Injected noise
  • 0.000 mV0.967
  • 0.025 mV0.944
  • 0.050 mV0.889
  • 0.100 mV0.814
  • 0.200 mV0.678

Synthetic noise is a crude proxy for real motion artifact, which is correlated across leads and structured in time rather than Gaussian.

Acute infarction versus healed scar
  • unknown84.5%
  • Stadium II-III85.4%
  • Stadium III95.9%
  • Stadium I92.9%
  • Stadium II100.0%

Caveat: Stage is recorded for only part of the cohort and the acute-stage subgroup is small, so these are indicative rates and not precise subgroup estimates. Wide uncertainty on any group with n below roughly 30.

Recall on MI-positive test records, grouped by PTB-XL's recorded infarction stage. Stadium I is acute, Stadium III is old. A model detecting acute events specifically would score Stadium I far above Stadium III. Similar rates across stages indicate the model is substantially recognising established infarct morphology, including long-healed scar.

Median latency
0.78 ms
p95 latency
0.87 ms
Parameters
411,329
Device
CPU

Raw artifacts in artifacts/model/. Full method, limitations, and bias risks in the transparency page and MODEL_CARD.md.