HeartBeacon

The expensive part of heart disease
is finding it late.

Cardiovascular disease is the largest single cost centre in American healthcare and the leading cause of death. It is also, by the WHO’s assessment, largely preventable. The gap between those two facts is where a screening model earns its keep.

That third number is the one that matters commercially. The WHO notes that “most cardiovascular diseases can be prevented by addressing behavioural and environmental risk factors” — tobacco, diet and obesity, physical inactivity, alcohol. Those are exactly the variables the CDC survey captures, and exactly what this model consumes.

So the model does not compete with an ECG or a lipid panel. It runs upstream of them, on data a health plan or employer already holds, at effectively zero marginal cost per person. That changes who can deploy it: not just hospitals, but payers, employer wellness programmes and public health agencies screening whole populations to decide who gets the expensive test.

Screening is a volume problem
before it is a modelling problem.

Cardiovascular disease is largely preventable when caught early, but screening every adult in a population is not economically possible. The question is therefore not who has heart disease — it is who is worth a clinician's time.

The dataset is the CDC's annual health survey: 319,795 respondents across 18 variables, mixing behavioural, demographic and self-reported clinical signals — BMI, smoking and alcohol use, stroke history, days of poor physical and mental health, difficulty walking, age band, diabetes status, physical activity, general health rating and sleep duration.

That makes it a triage dataset rather than a diagnostic one. Nothing here is a blood panel or an ECG. The realistic goal is to narrow a population down to a group worth testing properly, which sets the bar for what a useful model looks like.

The metric everyone reaches for
was actively misleading.

The target splits 10.68 to 1 — roughly 92% of respondents report no heart disease. That ratio has an uncomfortable consequence:

So accuracy was ruled out as a selection metric at the start. The metric that actually matters for screening is recall on the positive class: of the people who do have heart disease, how many did we flag? A missed case walks out of the clinic undiagnosed. A false alarm costs one follow-up test. Those are not symmetric errors, and the metric should not pretend they are.

The imbalance, visually. Across every continuous feature, the positive class (orange) is dwarfed by the negative class (blue). The distributions also overlap heavily — no single variable separates the classes cleanly, which rules out a simple threshold rule.

Fix the data, then fix the balance,
then compare honestly.

The work split into three stages. Each one was constrained by the same principle: do not let a preprocessing choice quietly inflate the final numbers.

Why outliers got per-feature treatment. BMI carries extreme high-side outliers that are implausible as measurements and were removed. Physical and mental health days cluster at zero with a long tail to thirty — that shape is the real distribution, so those were transformed instead.

Accuracy and recall pointed
at opposite models.

Random Forest won on accuracy by a wide margin and lost on the only metric that mattered. Plotted side by side, the trade-off is unambiguous:

Random under-sampling produced the strongest positive-class recall across model families, which is consistent with what it does: it removes the majority-class mass that was drowning the signal, at the cost of throwing away data. Here that trade was worth taking.

The honest caveat: F1 on the positive class sits at 0.33, which means precision is low — most flagged patients will not have heart disease. For a triage step that routes people to a cheap follow-up test, that is an acceptable cost. For anything that drives treatment, it is not. The model is fit for the first job and unfit for the second.

Choose the metric first,
and the model follows.

Logistic Regression with random under-sampling was selected — the lowest-accuracy model of the three, and the only one that did the job it was built for.

Choosing it also meant accepting a simpler model than the ensemble, which turned out to be an advantage rather than a concession. Logistic Regression exposes its coefficients, so a clinician can see which factors pushed a patient over the line. On a screening tool that someone has to trust and act on, an explainable model beats an opaque one with better paper metrics.

The underlying population. BMI is near-normal with a right skew; the health-day variables are zero-inflated with a spike at thirty; sleep time concentrates tightly between six and eight hours. These shapes drove both the transformation choices and the decision to standardise before modelling.

What it would take
to put this in front of patients.

A model that selects well in a notebook is perhaps a third of the way to something a health plan can actually run. The remaining work is mostly not modelling.

Where this falls short.

The model answers the question it was given. Several things would need to change before it could answer a harder one — and each of them taught me something I now reach for by default.