Naslov
A Supervised Classification Approach to Predicting Knee Pain Improvement in Osteoarthritis Patients

Autori
Banisakher, Deya ; Rishe, Naphtali ; Finlayson, Mark ; Marinovic, Ivanka

Vrsta, podvrsta i kategorija rada
Sažeci sa skupova, sažetak, znanstveni

Izvornik
Proceedings of the 30th International FLAIRS Conference / - , 2017, 774-774

Skup
30th International FLAIRS Conference

Mjesto i datum
Marco Island (FL), Sjedinjene Američke Države, 22.05.2017. - 24.05.2017

Vrsta sudjelovanja
Poster

Vrsta recenzije
Recenziran

Ključne riječi
Strojno učenje, automatsko klasificiranje, osteoartritis, predviđanje boli, projekcija ishoda, Ponavljajuće neuronske mreže
(Machine Learning, Automatic Classification, Osteoarthritis, Pain Prediction, Outcome Projection, Recurrent Neural Networks)

Sažetak
The data-driven prediction of an individual patient’s response to particular treatments is a major goal of precision medicine. We are tackling this problem in the domain of chronic diseases, specifically Osteoarthritis (OA), using the Osteoarthritis Initiative (OAI) dataset, which comprises OA-related medical records for more than 4000 patients over 10 consecutive years. Using these data, we have developed three new supervised machine learning classifiers that can determine at better-than-baseline rates as to whether osteoarthritis patients are experiencing improved, unchanged, or worsened pain relative to their previous assessment. Such a classifier is a necessary first step to predicting longer-term treatment outcomes. We use the standard Knee Injury and Osteoarthritis Outcome Score (KOOS) as labels, and train our classifiers on a set of easily observable features capturing demographics, related injuries, therapies (excluding medications), overall measures of pain, and measures of both physical activity and rest required during such activity. We trained three types of classifiers (Support Vector Machine, Random Forest, and a Multi-layer Neural Network), and all classifiers performed at better-than-baseline rates (baseline most- frequent-class gives 0.4 F1), with the neural network performing the best with over 0.7 F1. We further analyze which features are most predictive (particularly types and intensity ofrimarily walking and stand-sit activities coupled with the amount of time spent performing them), and identify several promising next steps for investigation, including integrating medications into the feature set.

Izvorni jezik
Engleski

Znanstvena područja
Interdisciplinarne biotehničke znanosti, Biotehnologija u biomedicini (prirodno područje, biomedicina i zdravstvo, biotehničko područje)

POVEZANOST RADA