Intrinsic Interpretability

The model's learned parameters are the explanation. In linear regression, each coefficient directly quantifies the expected change in the output for a one-unit increase in the corresponding biomarker.

• Linear Regression: A coefficient of +0.8 for LDL cholesterol means a direct, proportional increase in cardiovascular risk score.
• Logistic Regression: Weights represent the change in log-odds, which can be exponentiated to odds ratios for clinical interpretability.
• No Black Box: There is no secondary approximation or surrogate model required to understand the decision function.

Decision trees make predictions by applying a sequence of binary splits on individual features. This creates a human-traceable logic flow from root to leaf.

• Clinical Analogy: Mimics a differential diagnosis flowchart used by clinicians.
• Rule Extraction: Each path can be written as a conjunctive rule (e.g., IF Glucose > 180 AND BMI > 30 THEN High Risk).
• Global Transparency: The entire model's logic is visible at once, not just local approximations around a single prediction.

Generalized Additive Models (GAMs) constrain the model to learn a separate smooth function for each feature, which are then summed. This eliminates feature interactions from the core structure.

• Shape Functions: Each feature's effect is visualized as a curve, showing exactly how the risk changes across the feature's range.
• Modular Audit: A clinician can inspect the effect of 'Tumor Size' in isolation, confident that its contribution is not entangled with 'Patient Age'.
• Example: An EBM (Explainable Boosting Machine) learns these graphs automatically, providing state-of-the-art accuracy with glass-box transparency.

Domain knowledge can be encoded directly into the model architecture by forcing a monotonic relationship between a feature and the output. This guarantees logically consistent predictions.

• Guaranteed Behavior: A model predicting disease severity can be constrained so that increasing 'Tumor Grade' never decreases the risk score.
• Regulatory Alignment: Satisfies clinical intuition and prevents nonsensical predictions that erode physician trust.
• Implementation: Achieved in gradient-boosted trees like XGBoost and LightGBM via monotone constraints on specific features.

Intrinsically interpretable models often incorporate built-in mechanisms to zero out irrelevant features, producing a parsimonious model that focuses on a small set of high-impact biomarkers.

• LASSO Regression: Applies an L1 penalty to shrink coefficients of non-predictive features exactly to zero.
• Clinical Utility: A model that uses only 5 out of 10,000 genes is inherently more understandable and cheaper to deploy in a lab.
• Reduced Overfitting: Sparse models are less likely to learn noise, improving generalization to new patient cohorts.

While deep learning models are often opaque, architectures with built-in attention mechanisms can provide intrinsic rationales by design. The model learns to output both a prediction and the evidence it used.

• Hard Attention: The model explicitly selects a subset of input tokens (e.g., specific words in a clinical note) as its justification.
• Rationalizing Neural Networks: A generator module extracts a rationale, and a separate predictor makes the classification only from that rationale.
• Auditability: The rationale can be directly evaluated for clinical coherence, making the model's reasoning process falsifiable.

The foundational interpretable model where the predicted outcome is a weighted sum of input features. Each coefficient directly represents the expected change in the target variable for a one-unit increase in the corresponding feature, holding all others constant.

• Coefficient Sign: Indicates direction of relationship (positive/negative)
• Coefficient Magnitude: Indicates feature importance strength
• P-values: Quantify statistical significance of each feature
• Confidence Intervals: Provide a range of plausible true coefficient values

Clinical Example: Predicting patient blood pressure where a coefficient of +2.3 for sodium intake means each additional gram of daily sodium is associated with a 2.3 mmHg increase in systolic pressure, all else equal.

A classification model that estimates the probability of a binary outcome using a linear combination of features passed through a sigmoid function. Despite the non-linear transformation, the relationship between features and the log-odds of the outcome remains strictly linear and additive.

• Odds Ratios: Exponentiated coefficients quantify multiplicative change in odds per unit feature increase
• Decision Boundary: A single hyperplane separates classes in feature space
• Feature Contribution: The product of coefficient and feature value gives the log-odds contribution

Diagnostic Example: A model predicting 30-day hospital readmission where an odds ratio of 1.4 for prior admissions means each previous hospitalization increases readmission odds by 40%.

A hierarchical model that recursively partitions data using simple if-then rules on individual features. The entire decision process can be visualized as a flowchart, making it one of the most transparent model classes for clinical and regulatory applications.

• Root-to-Leaf Paths: Each path represents a complete decision rule
• Split Criteria: Gini impurity or entropy thresholds are directly inspectable
• Feature Thresholds: Exact cutoff values (e.g., glucose > 126 mg/dL) are explicit
• Leaf Nodes: Contain the predicted class distribution for that subgroup

Clinical Example: A diabetes screening tree where the first split on HbA1c ≥ 6.5% immediately separates high-risk patients, with subsequent splits on BMI and age refining the stratification.

An extension of linear models where the outcome is the sum of smooth, non-linear functions of individual features. GAMs preserve additivity—meaning there are no interactions between features—allowing each feature's effect to be visualized as a distinct shape function.

• Shape Functions: Plots showing how the predicted outcome changes across a feature's range
• Additive Structure: Total prediction = intercept + f₁(x₁) + f₂(x₂) + ... + fₚ(xₚ)
• Smoothness Penalties: Control the wiggliness of each shape function
• Degrees of Freedom: Quantify the complexity of each feature's contribution

Biomarker Example: Modeling disease risk where the shape function for age shows a gradual increase until 60, then a steep acceleration, while the function for biomarker concentration shows a threshold effect above 50 ng/mL.

Models that learn a set of explicit if-then rules directly from data. Unlike decision trees, rule sets are not constrained to a hierarchical structure—each rule is an independent conjunctive condition that can be inspected, modified, or removed without affecting others.

• Rule Antecedents: The conjunction of conditions (e.g., age > 50 AND smoker = true)
• Rule Consequent: The predicted class or outcome
• Rule Support: Number of training instances covered by the rule
• Rule Confidence: Accuracy of the rule on its covered instances

Clinical Example: A sepsis early-warning system with rules like "IF heart_rate > 90 AND respiratory_rate > 20 AND temperature > 38°C THEN flag high-risk," each independently auditable by clinicians.

Simplified linear models where coefficients are rounded to small integers, and predictions are made by summing points assigned to each feature condition. These are the most human-computable models, designed for rapid bedside calculation without computational assistance.

• Point Assignments: Integer weights for each feature bin or condition
• Total Score: Sum of points across all features
• Risk Strata: Predefined score ranges mapped to risk categories
• Transparent Thresholds: Exact score cutoffs for clinical action

Clinical Example: The APACHE II severity score where 12 physiological variables each contribute weighted points, and the total score directly maps to mortality risk—fully computable by hand at the bedside.