Can routine measurements screen for heart disease?

Coronary artery disease is confirmed by angiography — an invasive procedure. Before it is ordered, a clinician already has a patient’s age, blood pressure, cholesterol, resting ECG, and how they responded to an exercise test.

How much of the eventual diagnosis is already visible in those routine measurements?

This matters because the answer decides what a model like this could ever be for. If routine measurements carry most of the signal, a model built on them could help prioritise who gets referred. If they do not, then any model that scores well is quietly relying on tests that are only ordered once disease is already suspected — and it is answering a question that has already been answered.

The dataset here contains both kinds of feature, which makes the distinction testable rather than rhetorical.

What this study cannot answer

It cannot say whether any measurement causes heart disease. It cannot establish what would happen to a patient outside this sample. And it cannot produce a diagnosis — the label being predicted is the outcome of an angiogram that has already happened.

303 patients referred for cardiac evaluation at the Cleveland Clinic Foundation in the late 1980s. Thirteen measurements each, plus an angiographic diagnosis.

The full feature dictionary, licence, checksum and collection limitations are on the dataset page.

The leakage audit

Every supervised study here runs a leakage audit before any model is fitted, and training aborts if a feature correlates 0.90 or above with the target. Nothing did — the strongest was thal at +0.538. No duplicate rows, no constant columns, no entity identifier that could span a split, no temporal ordering.

But the audit surfaced something the correlation check alone would have missed.

ca (vessels coloured by fluoroscopy) and thal (thallium stress test) are results of invasive follow-up tests. They precede the diagnosis label, so they are not leakage in the strict sense. They are also not available at the moment a screening decision is made — they are ordered because disease is already suspected. ca in particular is close to a direct measurement of the thing being predicted.

A model using them does not screen. It reports what the catheterisation found.

So the study trains two variants: one with everything, and a screening variant with ca and thal dropped. The screening variant is the one served below.

Class balance. 164 patients without angiographic disease, 139 with — a 45.9% positive rate. Close to balanced, which is unusual for a clinical dataset and is a direct consequence of this being a referred population rather than a screened one. It means accuracy is less actively misleading here than it would be on a rare outcome, though it is still not the metric that carries the decision.

Missingness. Six values across 303 rows, confined to ca (4) and thal (2) — the two invasive features, which is itself consistent with those tests not always having been performed.

Strongest associations with the diagnosis, on the training split:

The two invasive tests top the list, which is exactly what makes them tempting and exactly why they were examined rather than simply used. Below them, the routine exercise-test measurements — maximum heart rate, chest pain type, exercise-induced angina — carry associations of similar magnitude.

Supervised binary classification. The unit of observation is one patient at one evaluation.

Target. num records 0 for no disease and 1–4 for increasing angiographic severity. The study binarises to presence versus absence. This is the conventional framing for this dataset, and it discards real information: a borderline case and a severe case become the same label. That is a choice the study made, not a property of the data.

Validation. A stratified 75/25 split with a fixed seed, giving 227 training and 76 test rows. Stratified because 303 rows are few enough that an unstratified split can shift test-set prevalence enough to move recall by several points on its own.

Model selection used 5-fold cross-validated recall on the training split only. The test split was evaluated once, at the end. Repeatedly consulting a 76-row test set while choosing between models would turn it into a second training set.

The metric that decides. Recall on the positive class. In a screening frame, a missed case costs more than an additional referral — so a model that trades precision for recall is making the right trade, within reason. Accuracy is reported, never headlined.

All preprocessing lives inside the serialised pipeline, so the model served below cannot apply a different transformation from the one it was trained with. This is not a stylistic preference: a preprocessing step applied separately at inference time is one of the few bugs that produces plausible wrong answers rather than an error.

Imputation statistics are fitted on the training split only. Fitting them on the full dataset before splitting is the most common quiet form of leakage — the test rows contribute to a median that is then used to fill training rows.

Unknown categories are ignored rather than raising, because a visitor to the lab below can submit a combination that 227 training rows never contained, and a server error is the wrong answer to that. The cost is recorded in the limitations: such an input is extrapolated silently.

Majority-class baseline. Predicts ‘no disease’ for everyone. It exists because without it, no other number means anything. It achieves 53.9% accuracy and 0.000 recall — a clean illustration of why accuracy was not chosen as the headline metric.

Logistic regression (C=1.0, class_weight='balanced'). Considered because its coefficients are directly inspectable, it is well suited to 227 rows, and it produces probability estimates natively. Assumes a linear relationship between features and log-odds — a real constraint, taken deliberately.

Random forest (300 trees, min_samples_leaf=3, balanced). Considered because it captures interactions and non-linearities logistic regression cannot. Risks overfitting a small sample and is harder to explain.

Both use balanced class weights so the slightly smaller positive class is not under-weighted. This has a consequence that matters later: balanced weighting deliberately distorts predicted probabilities away from the true base rate, which is why an explicit calibration step follows.

On the held-out test split of 76 patients. Intervals are 95% percentile bootstrap, 2000 resamples.

The comparison

See the table below. Then read this, because the table is easy to misread:

None of these models is distinguishable from the others. Logistic regression scores 0.886 recall and the random forest 0.857, but the intervals are [0.7778, 0.9737] and [0.7333, 0.9677] — they overlap almost entirely. With 76 test rows, a single patient moving from one side of the threshold to the other shifts recall by more than a point.

What the numbers do establish is that every candidate is far better than the baseline, which is a much weaker claim than a leaderboard implies, and the honest one.

Why logistic regression was selected

Not because it scored highest. It was selected because when scores are indistinguishable, the tiebreakers are interpretability, robustness on a small sample, and cost to serve — and it wins all three. Choosing the top row of a table whose rows are statistically identical is how a model gets picked for noise.

Confusion matrix

Of 35 patients who had disease, the model identified 30 and missed 5. Of 41 who did not, it correctly cleared 36 and flagged 5 who were disease-free.

In a screening frame those five false positives are the acceptable cost of the trade. The five false negatives are the number that matters, and the limitations section says who they tend to be.

Did dropping the invasive tests hurt?

Measurably, no. The full-feature variant reached 0.857 recall; the screening variant reached the same, with overlapping intervals. On this sample the routine measurements carry the available signal. That result is what justifies serving the screening model — it gives up nothing detectable and answers a question worth asking.

Logistic regression coefficients are directly inspectable, which is most of why it was chosen. The served model reports the inputs it weighs most heavily alongside each prediction.

One caveat is stated wherever those weights appear: they are global importances for the fitted model, not an attribution for the individual prediction on screen. A feature with a large coefficient influenced the model’s overall behaviour; it did not necessarily drive this particular result. Per-prediction attribution would need SHAP or a comparable method, which is not justified for a linear model on eleven features.

On calibration

The model is explicitly calibrated using Platt scaling, and the reason is specific: class_weight='balanced' deliberately shifts the decision boundary toward the positive class, which pushes raw predicted probabilities away from the population base rate. Without correction the numbers would rank cases correctly while being wrong as probabilities.

Reported honestly: the calibrated Brier score (0.117) is not better than the uncalibrated logistic regression’s (0.111). At 76 test rows that difference is noise. The calibration step is kept because it is the theoretically correct response to the class weighting, not because the evidence shows it helped.

And the probability remains calibrated to a referred population with 45.9% prevalence — not to a general one.

The screening model, running live. Every input is a measurement available before any invasive test — that is the whole point of this variant.

The output is a class label and a calibrated probability, the decision threshold that produced it, and the model’s semantic version. Read the probability as “among referred cardiac patients in this 1988 sample with similar measurements, roughly this proportion had angiographic disease.” It is not a statement about you, and the ranges below are the ranges the model actually saw.

What was observed. Among 303 referred cardiac patients, routine pre-angiographic measurements — chest pain type, exercise-induced angina, maximum heart rate, ST depression — are associated with angiographic disease strongly enough that a linear model separates the classes well on held-out data.

What the model does. It identifies 86% of positive cases in the test split, with a 95% interval of [0.7333, 0.9677], at 86% precision.

What is inferred. That the invasive tests were not carrying the signal alone. Removing both cost nothing measurable, which suggests the routine measurements already contain most of what this dataset has to offer.

What cannot be concluded.

• That any measurement causes heart disease. This is an observational association, not a causal design.
• That logistic regression is better than a random forest for this problem. The confidence intervals say those models are indistinguishable here.
• That these numbers would hold in any other population, era, or care setting.
• That a probability from this model applies to an individual outside the referred population it was fitted on.

The most useful result is not the recall figure. It is that examining when a feature becomes available changed which model got deployed — and cost nothing to do.

Ordered by how much each should change your reading of the result.

The two that matter most: the test set has 76 rows, so every interval here is wide enough that model comparisons do not survive it; and this is a referral population, so the probabilities are conditioned on someone having already decided this patient warranted cardiac evaluation.

Sources are listed below.

Recall — of everyone who actually had the condition, the share the model found. It is the metric that carries the decision here because a missed case costs more than an extra referral.

Precision — of everyone the model flagged, the share who actually had it. Recall and precision trade against each other; moving the threshold buys one with the other.

Calibration — whether a predicted probability of 0.7 corresponds to the event happening about 70% of the time. A model can rank cases perfectly and still be badly calibrated, which is why ranking metrics like AUC and calibration metrics like the Brier score are both reported.

Class weighting — telling the model to treat errors on the smaller class as more costly. It improves recall and, as a side effect, de-calibrates the raw probabilities, which is why a calibration step follows it here.

Target leakage — a feature that encodes the answer. The strict form is caught by a correlation check. The form that caught this study is subtler: a feature legitimately measured before the label, but not available at the moment the prediction would need to be made.

Bootstrap confidence interval — resampling the test set with replacement to see how much a metric moves. On 76 rows it moves a great deal, which is the single most important thing to know about every number on this page.