Appointment No-Show Risk Model

A prediction model built on past booking history that sorts tomorrow’s appointments into risk bands, each with a short reason, so reminder calls go where they matter.

Machine learning that predicts which appointments are likely to be missed, so staff confirm those first.

Reminder calls went out in schedule order, so staff spent their time on patients who always attend while the slots that went empty were never confirmed.

The clinic had years of booking history in MySQL but no way to use it to prioritize the day.

Reminder calls went out in schedule order. Staff spent the morning on patients who always attend, and the slots that went empty were the ones nobody had time to confirm.

• •Score tomorrow’s appointments by the chance they will be missed.
• •Give staff a short reason for each score, not just a number.
• •Fit into the existing Laravel scheduler without changing how staff work.

I framed the problem, built the feature pipeline and model in Python with scikit-learn and Pandas, designed the evaluation, and integrated the scores into the Laravel scheduler through Redis.

A nightly Python job builds features from past bookings in MySQL, scores tomorrow’s appointments with a regularized logistic regression model, and the Laravel scheduler reads the risk band from Redis to sort the confirmation queue.

1. 01

   Feature build (Pandas)

   A nightly job reads past bookings from MySQL and builds features known before the visit: lead time, weekday, slot hour, visit type, and prior attendance history.

2. 02

   Model (scikit-learn)

   A pipeline of scaling, one-hot encoding and regularized logistic regression with balanced class weights.

3. 03

   Risk bands

   Probabilities are converted into three bands with thresholds chosen on validation data, plus the top contributing features from the coefficients.

4. 04

   Laravel integration

   Bands are written to Redis; the scheduler sorts the confirmation queue by band.

• •Applies supervised classification from the Supervised Machine Learning: Regression and Classification course: logistic regression, feature scaling, regularization and decision thresholds.
• •Builds the end-to-end workflow in Python taught in Machine Learning with Python, from Pandas features to a scikit-learn pipeline.
• •Explainable output: every band comes with the features that pushed it up.

pipeline = Pipeline([
    ("prep", ColumnTransformer([
        ("num", StandardScaler(), ["lead_days", "prior_no_shows", "prior_visits"]),
        ("cat", OneHotEncoder(handle_unknown="ignore"), ["weekday", "slot_hour", "visit_type"]),
    ])),
    ("clf", LogisticRegression(C=0.5, class_weight="balanced", max_iter=1000)),
])
pipeline.fit(X_train, y_train)
bands = np.digitize(pipeline.predict_proba(X_tomorrow)[:, 1], [0.35, 0.6])

1. 01

   Missed visits are the minority class, so the model is judged on precision and recall at the chosen threshold, not on raw accuracy.

   Balanced class weights and evaluation on precision and recall at the chosen threshold, with the threshold selected on a validation set.

2. 02

   Features had to come only from data known before the visit, so nothing from the appointment outcome leaks into training.

   Features are restricted to what is known at booking time, and the split is by date so no future information leaks into training.

3. 03

   Staff need a reason, not a score, so each band ships with the top contributing features from the model coefficients.

   Coefficient-based explanations attached to each band, shown in the queue next to the appointment.

Security

The model reads booking metadata, not clinical notes. Scores are stored as bands in Redis with the same access rules as the schedule.

Speed & growth

Scoring runs nightly in batch, so the scheduler reads precomputed bands with no model call at request time.

Held-out evaluation on a later time window, checks for leakage in feature generation, and a review of explanations with front-desk staff before rollout.

These are design targets and approach choices for the system, not reported production figures.