Leveraging automated machine learning to benchmark, deconstruct, and compare frailty indices for predicting adverse spinal surgery outcomes

Figures

• 

  Fig. 1. Stacked bar plots of proportions and counts for each HFRS (left) and eFI (right) frailty level across outcome events.

• 

  Fig. 2. Shapley value–based feature importance for top predictors in final models across all outcomes. Panels (A, C, and E) display mean absolute Shapley values, indicating global feature importance for DD, LOS, and RE, respectively. Panels (B, D, and F) provide corresponding local explanation plots that show the direction and magnitude of each feature’s effect on prediction. Red denotes high feature values (for ordinal features, e.g., Hemoglobin) or presence (for binary features, e.g., M41), while blue denotes low values or absence. For DD (B), negative SHAP values correspond to increased likelihood of DD_NH; for LOS (D) and RE (F), positive SHAP values indicate a higher probability of LOS>6d and RE90, respectively. For example, low Hemoglobin values (blue), are associated with adverse outcomes in both DD_NH (B) and LOS>6d (D) despite being on opposite sides of 0. Feature names followed by asterisks (*) are HFRS components while features names followed by daggers (†) are eFI components.

• 

  Fig. 3. Annotated heatmaps of case ratio deviations across top two feature interactions by outcome. Each row represents one clinical outcome: DD (top), LOS (middle), and RE (bottom). Tile shading indicates the deviation of the subgroup-specific case ratio from the global outcome case ratio. Tiles shaded in orange represent negative (adverse) deviations. For DD, higher case ratios (more home discharges) are favorable and are shaded in blue. In LOS and RE, higher case ratios indicate worse outcomes (shaded in orange). Each tile is annotated with the case ratio (R), subgroup size (n), and FDR-adjusted p-value (‘*’ indicates FDR < 0.05).