Concept Drift

An abrupt, discontinuous shift in the data distribution caused by an external event. In a clinical context, this is often triggered by an instantaneous operational change.

• Mechanism: A new diagnostic device is deployed, instantly changing the scale and distribution of lab values.
• Example: A hospital switches troponin assay vendors, causing a sudden jump in reported values that breaks an acute myocardial infarction predictor.
• Impact: Immediate and catastrophic model failure if not detected in real-time.

A gradual, continuous evolution of the target concept over time. This is the most common form of drift in healthcare, reflecting slow changes in clinical practice.

• Mechanism: Progressive adoption of new treatment guidelines that slowly alters patient outcomes.
• Example: Over 5 years, a sepsis predictor degrades as standard-of-care antibiotics shift from broad-spectrum monotherapy to combination therapy, changing the mortality signature.
• Detection: Requires monitoring statistical divergence metrics like Population Stability Index (PSI) over rolling windows.

Cyclical or seasonal patterns where the data distribution oscillates predictably between two or more states. The model must generalize across these recurring contexts.

• Mechanism: Temporal patterns that repeat on a known cadence, such as time-of-day, day-of-week, or seasonal variation.
• Example: A pediatric RSV early warning score that performs well in winter but drifts in summer due to low disease prevalence and different patient demographics.
• Mitigation: Contextual features (e.g., month, season) or separate seasonal models.

A shift in the distribution of the input features (covariate shift) that does not necessarily change the decision boundary but triggers a false alarm in monitoring systems.

• Mechanism: A change in the patient population that alters the input data histogram without invalidating the underlying clinical relationship.
• Example: A new primary care clinic opens, shifting the average patient age lower. The model's input distribution changes, but the relationship between age and disease risk remains stable.
• Importance: Distinguishing virtual drift from real concept drift prevents unnecessary and costly model retraining.

A specific subtype of drift where the statistical properties of individual predictor variables change, potentially breaking feature engineering pipelines.

• Mechanism: Changes in measurement units, coding practices, or data quality for a specific input.
• Example: A clinical NLP model drifts when a hospital system updates its ICD-10-CM coding guidelines, changing the distribution of extracted comorbidity concepts.
• Monitoring: Univariate statistical tests (Kolmogorov-Smirnov) on each feature against a reference baseline.

A change in the definition, interpretation, or recording of the target variable itself, directly invalidating the ground truth.

• Mechanism: Administrative or clinical redefinition of the outcome being predicted.
• Example: A model predicting 'hospital-acquired pressure injury' drifts after the health system updates its staging criteria from the NPUAP 2016 to the 2019 guidelines, changing what constitutes a positive case.
• Challenge: Cannot be detected by input monitoring alone; requires continuous auditing of outcome distributions and clinical definitions.

A change in the statistical distribution of the input features (independent variables) over time, while the relationship to the target variable remains constant. Unlike concept drift, the decision boundary is still valid, but the model encounters unfamiliar data.

• Example: A new pulse oximeter vendor changes the waveform data format fed to a sepsis predictor.
• Detection: Population Stability Index (PSI), Kolmogorov-Smirnov tests on feature distributions.
• Remediation: Model recalibration or input preprocessing adaptation.

A shift in the marginal distribution of the target variable P(y) itself, independent of the input features. The prior probability of an outcome changes, but the conditional relationship P(y|X) may remain stable.

• Example: A hospital system acquires a cardiac specialty center, dramatically increasing the prevalence of heart failure diagnoses in the patient population.
• Impact: A well-calibrated model will appear miscalibrated because the base rate has shifted.
• Detection: Monitoring the mean of predicted probabilities against observed event rates.

A specific type of data drift where the distribution of input features P(X) changes, but the conditional distribution of the target given the features P(y|X) remains identical. This is a feature-space only problem.

• Example: During COVID-19, the age distribution of hospitalized patients shifted younger, changing the input demographics for a length-of-stay predictor.
• Key Distinction: The clinical relationship between age and length of stay didn't change, but the model sees a different population.
• Technique: Importance-weighted retraining to correct for the sampling bias.

The gradual erosion of a model's predictive performance over time due to any combination of drift types, data staleness, or environmental changes. This is the observable symptom; drift is the underlying cause.

• Monitoring: Tracking AUC, F1-score, or Brier score degradation on a holdout set.
• Clinical Risk: A decaying sepsis predictor may silently miss cases, creating a false sense of security.
• Mitigation: Automated retraining pipelines triggered by performance thresholds.

The specific degradation of a model's probabilistic accuracy over time. A predicted 20% risk no longer corresponds to a 20% observed event rate. The model's ranking ability (discrimination) may remain intact, but its absolute risk estimates become unreliable.

• Clinical Significance: Critical for decision support tools where thresholds trigger interventions.
• Detection: Expected Calibration Error (ECE) and reliability diagrams.
• Remediation: Platt scaling or isotonic regression on a recent calibration set.

An abrupt, discontinuous change in the data-generating process, as opposed to gradual or incremental drift. Often triggered by an external shock or system change.

• Example: A new clinical guideline redefines the diagnostic criteria for sepsis, instantly invalidating the labels used to train an existing predictor.
• Detection: Change-point detection algorithms and real-time monitoring dashboards.
• Response: Requires immediate model freeze and emergency retraining, not gradual adaptation.