@Rijnbeek, @anthonysena and I had a great opportunity to take part in @Daniel_Prieto’s Oxford Summer School course on Real-World Data Epidemiology this week. As part of our segment of the training, the class identified a series of important clinical problems that could be informed through patient-level prediction, and chose one specific question to design and implement as a group.
The team settled on the following prediction problem:
Amongst patients initiating bisphosphonates after diagnosis of osteoporosis and after 40 years of age, which patients will have a hip fracture event within the 2 years following diagnosis?
We used ATLAS to define our target and outcome cohorts, then designed an analysis script using the PatientLevelPrediction package to learn models against any OMOP CDM-compliant database. We then used one of the newest features @jennareps created to externally validate the model on other data.
The preliminary results, generated during the course, were extremely promising. We trained a LASSO regression against a large US claims database, which yielded a XX-variable model that achieved a AUC=0.82 on the hold-out test set. The model was then externally validated against 5 different databases from UK, Japan, Germany, and other US claims/EHR data all achieve AUC>0.70 when directly applying the original model.
Based on the encouraging feasibility work, @Daniel_Prieto, @Rijnbeek and I intend to work with the students in the class to take this research forward to a full publication. We want to make this an OHDSI network study to allow others to participate in the fun! We will post a protocol and study package in the near future, and invite all of you to externally validate our model and participate in the publication. If folks want to participate or have thoughts about this work, please join this discussion thread!
Thanks @Daniel_Prieto for a wonderful week in Oxford. Great job and thank you for your contributions to the OHDSI community!
. I think that interpretation of the outcome would be important.