Partial Dependence Plot (PDP)

A PDP isolates the relationship between a target feature and the predicted outcome by marginalizing out the influence of all other features. The algorithm averages the model's predictions over the empirical distribution of the complementary set, revealing whether the relationship is linear, monotonic, or more complex. This is critical for verifying that a diagnostic model has learned a clinically plausible risk curve, such as a monotonic increase in disease probability with rising biomarker concentration.

PDPs are a post-hoc, model-agnostic explanation method. They treat the predictive model as a black box, requiring only the ability to generate predictions from perturbed input data. This property makes them universally applicable across model architectures—from logistic regression and random forests to deep neural networks—allowing FDA submission teams to use a single, consistent interpretability technique across diverse diagnostic pipelines.

While one-way PDPs show the effect of a single feature, two-way PDPs visualize the joint effect of two features on the predicted outcome. Rendered as contour or surface plots, they expose interaction effects that are invisible in univariate analysis. For example, a two-way PDP can reveal that a genomic marker only increases risk when a specific proteomic value is also elevated, providing mechanistic insights into disease etiology.

The primary limitation of PDPs is the assumption that the target feature is independent of the complementary features. When features are highly correlated, the averaging process can produce unrealistic synthetic data points in low-density regions of the feature space, leading to misleading interpretations. For correlated biomarkers, Accumulated Local Effects (ALE) plots are a more robust alternative that avoids this extrapolation problem.

PDPs show the average marginal effect across all instances, which can mask heterogeneous effects. If a feature has a strong positive effect on half the population and a strong negative effect on the other half, the PDP may incorrectly show a flat, non-influential relationship. This is a critical consideration in patient stratification, where Individual Conditional Expectation (ICE) plots should be used alongside PDPs to reveal divergent response patterns.

PDPs naturally extend to categorical features by computing the average prediction for each discrete category level. For a biomarker like 'genetic mutation status' (Wild-type, Heterozygous, Homozygous), the PDP displays the model's expected output for each variant, providing a direct, clinically interpretable comparison of risk across genotypes. This is often more intuitive than numerical feature importance scores for communicating with clinical stakeholders.

Disaggregates a PDP by plotting the prediction curve for each individual instance rather than averaging. ICE plots reveal heterogeneous effects that PDPs can mask when relationships cancel out across subgroups.

• Each line represents one observation's response to feature changes
• Clustered ICE lines expose distinct behavioral cohorts
• Essential for detecting interaction effects hidden by averaging

A faster, unbiased alternative to PDPs when features are correlated. ALE plots isolate the local effect of a feature by computing differences in predictions within small intervals, then accumulating them.

• Avoids the extrapolation problem of PDPs in correlated regions
• Computationally more efficient than PDPs
• Centers effects at zero for easier comparison across features

Quantifies the strength of interaction between two features using the Friedman's H-statistic. While PDPs visualize marginal effects, the H-statistic measures how much of a model's prediction variance comes from joint feature effects.

• Ranges from 0 (no interaction) to 1 (pure interaction)
• Computed by comparing partial dependence of feature pairs against individual effects
• Guides which 2D PDPs are worth investigating

Measures a feature's global importance by randomly shuffling its values and observing the drop in model performance. Complements PDPs by ranking which features most deserve marginal effect analysis.

• Model-agnostic and simple to implement
• Captures both main effects and interactions
• Repeated permutations provide confidence intervals for importance scores

Extends beyond correlational PDPs by identifying features that are direct causes of outcomes using structural causal models. Essential for clinical diagnostics where understanding causation, not just association, determines treatment decisions.

• Uses do-calculus and causal graphs
• Distinguishes confounders from true drivers
• Critical for FDA submissions requiring mechanistic rationale