Lumbar spine problems occur widely in the U.S. Treatment of lumbar spine patients is generally a tedious and expensive process. In order to improve treatment efficiency and reduce unnecessary costs, Geisinger Health Systems has deployed a new protocol called ProvenCare Lumbar Spine. In this study we develop a machine learning approach to be employed as part of the ProvenCare protocol. We use graphical models to realize different treatment pathways and to estimate costs for patients with different demographics, severity levels and different treatment outcomes. We use existing medical records data accumulated over the years and combine it with domain knowledge provided by physicians to come up with a Bayesian network prediction model. We divide the model into three components and treat them separately, grouping certain related variables into relatively isolated dense subgraphs and by estimating the prior probabilities of respective state variables from data. First component is the modeling of the conservative management options for spine patients, where we predict the probability that a patient responds positively to either of three therapy types: i) oral medications, ii) physical therapy, iii) injection therapy. Second, for those failing conservative management, we model the pre-surgical tests conducted as part of the ProvenCare protocol to clinically optimize patient groups preceding a spine surgery. Third component is prediction of surgical outcomes and modeling of post-surgical procedures represented as random variables.
Our graphical model contains several random variables to represent features such as patient demographics, severity of diagnosed disease, comorbidities, current active diagnoses and active list of medications that the patient is currently on, to name a few. We took advantage of the conditional independence of variables modeled by means of Bayesian networks by making use of BayesiaLab’s graphical model editor and built-in inference/prediction algorithms. We have three targets to achieve: i) forward execution of the Bayesian network to predict the outcome of treatment protocols for different patients (patient groups), ii) backward execution of the model to determine the characteristics of the corresponding patient subpopulation for a predetermined surgical outcome, and iii) incorporation of cost information to estimate partial or complete costs of the procedures applied to patients.
This will be a specific use case of BayesiaLab for a medical application. In addition to employing the rich modeling and inference functionality of BayesiaLab, what is further interesting is that, we plan to show a workaround for learning with a mixed representation of disjoint feature subsets for the three components described above. To our knowledge, this is something not readily available via the data loading mechanism of BayesiaLab. We will also present quantitative results with real data from Geisinger, which is an exciting and distinctive dimension of this study.