MY DAUGHTER LOVES THE NEW PENS!- QUANTIFYING THE PATIENT EXPERIENCE WITH MACHINE READING AND APPLIED SEMANTIC COMPUTING
Author(s)
Bichteler A1, Collins BG2, Walter S1, Wendler K1, Kölling J1, Loonus Y1, Höwelkröger J1, Matheus C1, Jebbara S1, Hommel F1, Badmaeva E1, Verissimo S1, Mokbel B1, Cimiano P3, Hartung M1
1Semalytix, GmbH, Bielefeld, Germany, 2Boehringer Ingelheim, GmbH, Ingelheim, Germany, 3Bielefeld University, Bielefeld, Germany
Presentation Documents
OBJECTIVES. Real-world experience of disease treatment lies at the heart of patient centricity. Conventional methods of developing patient-reported outcomes (PROs) instruments and value assessments are often costly, burdensome, even impossible (e.g. in orphan diseases; pediatrics). Our goal was to generate patient insights from online forums on 1) Lupus Nephritis (LN), and 2) subcutaneous treatments in Crohn’s, with the confidence necessary for decision-making. METHODS. Machine reading analyzers “read” publicly available, anonymized forum posts: 22,500+ on lupus and 13,000+ on Crohn’s. Posts were split into sentences, and custom word embeddings and pharma-specific knowledge graphs represented the meaning necessary for entity and relation extraction. Decision makers identified relevant texts and supervised/unsupervised topics, e.g. sentiment, symptoms, and convenience. Quality was assessed by algorithmic confidence and expert evaluation. RESULTS. 7,562 lupus posts related to the LN subgroup. Pain and rash were discussed equally, but swelling 3.5 times as often in the LN vs. lupus-only group. Symptoms were differentiated by body part; rash and swelling in face, hand, and neck accounted for 53-69% of mentions. On the experience of subcutaneous devices, 3110 sentences related to syringe or auto-injector pens. Positivity for subcutaneous administration rose steadily 2008-2017; positivity for syringes surpassed that for pens in 2013. Discussions on convenience (52.4% among supervised topics) and pain (42.5% among unsupervised topics) revealed novel strategies for self-administration. Algorithmic performance in detecting sentiment amounted to 90% precision. Negative sentiment was the least precise, and false negatives (22%) exceeded false positives (10%). CONCLUSIONS. Machine reading technologies can quickly identify and quantify the patient experience where it is already abundant in social media and advocacy forums. Insights from focused research and discovered unknowns can inform decisions across the development and commercialization pipelines.
Conference/Value in Health Info
2019-11, ISPOR Europe 2019, Copenhagen, Denmark
Code
PMU128
Topic
Methodological & Statistical Research, Patient-Centered Research
Topic Subcategory
Artificial Intelligence, Machine Learning, Predictive Analytics, Instrument Development, Validation, & Translation, Patient-reported Outcomes & Quality of Life Outcomes
Disease
Gastrointestinal Disorders, Medical Devices, Multiple Diseases, Rare and Orphan Diseases