• Computer code • Learning model • Rules-based systems • Learns from examples • No flexibility • Inputs (Features) • Outputs (Labels) Machine vs Human
Machine learning Human learning
• Required more example to make • Small amounts of data general associations • Common sense • Can learn from massive amount of data • No common sense Supervised machine learning Application of Machine learning • Prognosis • Workflow • Parameter > Clinical outcomes • Predictive typing • Diagnosis • Automatic summarization • Suggest test • Surgery assistant • Raise awareness of diagnosis • Treatment • Recommend treatment • Compare treatment with model • Treatment based on expert/new clinical guideline/update paper • Identify pt. who may benefit from new drug/therapy under study Key challenges • Assembling representative data set • Ideal • Train model with data mostly resemble exact format expected during use • Data from trial • Need to have same specifications to be optimal • Limitations • Model not generalized to particular scenario • Overreliance > bias, decrease vigilance for errors • Need for prospective, real-world clinical evaluation • Retrospective study • Spinal epidural abscess age > 18 years Dx. By CT/MRI • Tx: Conservative (Abx +/- IVR drainage) • N: 472/1053 Tx conservatively Outcomes and variables • Primary outcomes: Failure of conservative Tx • Neurologic deterioration • Worsened back/radicular pain • Persistent symptoms > Change in Mx • Progression based on imaging > Change in Mx • Variables • Demographic, Signs and symptoms, duration of symptoms • Lab, Microbiology, Radiographic characteristic • Concurrent spinal/nonspinal infections Machine learning methods • Population: 80:20 split • Derivation (80%) • Use for machine training • Validation (20%) • Use for evaluating • 5 Machine learning methods • Elastic-net penalized logistic regression • Stochastic gradient boosting • Random forest • Support vector machine • Neural network Results • Variables include in final algorithm • Motor deficit • DM • Ventral component of abscess related to thecal sac • Hx of compression/pathologic fracture • Sensory dysfunction • Active malignancy • 3 or more spinal level • Best model: Elastic-net penalized logistic regression Discussion • Conservative in SEA • Normal neurological exam • Extensive panspinal infection • Complete paralysis for > 72 hr • Poor surgical candidacy • Refusal to Sx • Previous risk for failure • DM, Leukocytosis > 12500, H/C positive, CRP > 115.7 • Age > 65, DM, MRSA, Pretreatment motor deficit • Active malignancy, Fracture at affected level • Negative factor: Abscess at dorsal related to thecal sac https://sorg-apps.shinyapps.io/sean onop/