COMPARISON OF A BAYES-PRICE-LAPLACE CLINICAL LEARNING MACHINE TO CLINICAL PREDICTION RULES FOR PULMONARY EMBOLISM
To quantify and compare the performance of all published clinical prediction rules (CPRs) for pulmonary embolism (PE) to a Bayes-Price-La Place Clinical Learning Machine (BPLCLM).
Method(s):
The study population is 310 (65 PEs) consecutive patients referred for CT angiogram of the chest to rule out PE. All relevant clinical findings for each patient are entered into a tailored database that supports the development and evaluation of Bayes’ Rule parameters. The first 201 patients are used to determine the prior odds based upon the presenting findings for each diagnosis and the likelihood ratios for all clinical findings for each diagnosis. Predictor variables (PVs) that are utilized have a level of significance > .95, face validity, and enhance the ability to discriminate in the development set. PVs may be a single clinical finding or a group of clinical findings with either an “and” or “or” relationship; PVs may not be a clinical judgement. The sum of the probability for all diagnoses is proportionally constrained to sum to 1. The last 109 patients are used to calculate a P(PE) for each patient with each CPR and the BPLCLM. ROC curves are generated for each method.
Result(s):
| Clinical Prediction Method | ROC Area |
| Charlotte Score | .66 |
| Miniati Regression Equation | .74 |
| PERC Score | .61 |
| Revised Geneva Score | .59 |
| Simplified Revised Geneva Score | .55 |
| Simplified Wells Score | .60 |
| Wells & Charlotte Scores | .63 |
| Wells & PERC Scores | .62 |
| Wells Score | .61 |
| BPLCLM | .92 |
Conclusion(s): The BPLCLM has significantly more capacity to discriminate between patients with and without PE than any published CPRs. The enhanced performance of the BPLCLM is due to the adjustment of the prior odds based upon the presenting findings and utilization of all PVs that have been found to be relevant to the patient’s situation. If the BPLCLM function were to be incorporated into electronic medical record systems it would enhance the ecology of clinical decision making. Bayes rule is to clinical decision making, as the Pythagorean Theorem is to architecture and Newton’s Second Law of Motion is engineering; violation of the equation in their respective domains leads to falling buildings, crashing vehicles; and costly, risky, low-value health care.