Table 3.
Prediction ability of the reference and four machine-learning-based prediction models for top 1% or 10% healthcare cost spenders.
Outcome | c-statistics | P-valueb | Sensitivity | Specificity | PPV | NPV | PLR | NLR |
---|---|---|---|---|---|---|---|---|
The prediction model for top 1% healthcare cost spenders | ||||||||
Reference modela | 0.85 (0.82–0.87) | [Reference] | 0.71 (0.67–0.76) | 0.84 (0.83–0.84) | 0.04 (0.04–0.05) | 0.99 (0.99–0.99) | 4.4 (4.1–4.7) | 0.35 (0.29–0.41) |
Logistic regression with Lasso regularization | 0.86 (0.84–0.88) | 0.42 | 0.78 (0.73–0.82) | 0.78 (0.78–0.79) | 0.04 (0.03–0.04) | 0.99 (0.99–0.99) | 3.6 (3.4–3.8) | 0.29 (0.24–0.35) |
Random forest | 0.83 (0.80–0.85) | 0.26 | 0.66 (0.61–0.71) | 0.88 (0.87–0.88) | 0.05 (0.05–0.06) | 0.99 (0.99–0.99) | 5.4 (4.9–5.8) | 0.39 (0.34–0.45) |
Gradient-boosted decision tree | 0.85 (0.83–0.88) | 0.69 | 0.70 (0.65–0.74) | 0.87 (0.87–0.88) | 0.05 (0.05–0.06) | 0.99 (0.99–0.99) | 5.4 (5.0–5.8) | 0.35 (0.30–0.41) |
Deep neural network | 0.85 (0.82–0.87) | 0.91 | 0.74 (0.69–0.78) | 0.80 (0.80–0.80) | 0.04 (0.03–0.04) | 0.99 (0.99–0.99) | 3.7 (3.4–3.9) | 0.33 (0.28–0.39) |
The prediction model for top 10% healthcare cost spenders | ||||||||
Reference modela | 0.85 (0.85–0.86) | [Reference] | 0.74 (0.73–0.76) | 0.83 (0.83–0.84) | 0.33 (0.32–0.34) | 0.97 (0.97–0.97) | 4.4 (4.3–4.6) | 0.31 (0.29–0.33) |
Logistic regression with Lasso regularization | 0.85 (0.85–0.86) | 0.99 | 0.74 (0.73–0.76) | 0.83 (0.83–0.84) | 0.33 (0.32–0.34) | 0.97 (0.97–0.97) | 4.5 (4.3–4.6) | 0.31 (0.30–0.33) |
Random forest | 0.87 (0.86–0.88) | <0.001 | 0.75 (0.73–0.76) | 0.87 (0.87–0.88) | 0.39 (0.38–0.41) | 0.97 (0.97–0.97) | 5.8 (5.7–6.1) | 0.29 (0.27–0.31) |
Gradient-boosted decision tree | 0.88 (0.87–0.88) | <0.001 | 0.76 (0.75–0.77) | 0.87 (0.86–0.87) | 0.38 (0.37–0.40) | 0.97 (0.97–0.97) | 5.6 (5.4–5.8) | 0.28 (0.26–0.30) |
Deep neural network | 0.88 (0.87–0.88) | <0.001 | 0.75 (0.74–0.77) | 0.87 (0.87–0.88) | 0.39 (0.38–0.41) | 0.97 (0.97–0.97) | 5.8 (5.6–6.0) | 0.28 (0.27–0.30) |
PPV positive predictive value, NPV negative predictive value, PLR positive likelihood ratio, NLR negative likelihood ratio.
aWe used a non-penalized logistic regression model as the reference model.
bWe compared the area under the curve between each machine-learning-based prediction model and the logistic regression model (the reference model) using the DeLong’s test.