(108) Automated facial expression analysis can detect clinical pain in youth in the post-operative setting

Abstract

s The Journal of Pain S3 (108) Automated facial expression analysis can detect clinical pain in youth in the post-operative setting J Huang, K Craig, D Diaz, K Sikka, A Ahmed, L Terrones, G Littlewort, M Goodwin, and M Bartlett; University of California San Diego, La Jolla, CA Clinical pain assessment has relied on self-report and clinician observations. The former requires cognitive competence, social skills and candor and the latter requires clinician skill and substantial human resources. Failures to recognize pain and suboptimal pain assessment in youth have been documented, particularlywhenusing assessmentby proxy.Wepursued an alternative approachusing computer vision, pattern recognition and machine learning to provide automated, objectivepainassessment information in thepostoperative setting. Facial expressions have been recognized as sensitive and specific to pain and well-validated coding systems (FACS) are available. We utilized an automated FACS system (CERT) to measure facial expressions in clinical pain and no-pain situations to determine potential utility of CERT tomeasure clinical pain in children. Specifically, we video-recorded facial expressions within 24 hours of appendectomy in 40 otherwise healthy children and again after clinical resolution at a follow-up visit 20 (18,29) [median(interquartile range)] days later when no pain was reported. Measured youth were 13 (10,15) years old, 88% Hispanic, and 65% male. Recordings were analyzed using CERT, and mean measurements of facial actions associatedwith pain (AUs 4, 6, 7, 9, 10, 25, 43)were comparedwithin subjects over time using repeated-measure analyses. CERTwas able to detect differences in at-rest facial expressions during the immediate postoperative pain period (validated by self-reports of presence of pain) as compared to expressions measured during the no-pain period (AU4, P=0.006; AU7, P 18.61; p’s .335). The PDI and VAS showed mostly weak relationships with performances on sit-to-stand, stair climb and treadmill distance (r’s range: -.31 to .01; p’s >.06). Pain intensity during the functional tests (postminus pre-pain NRS) also showed weak relationships with performances on sit-to-stand, stair climb and treadmill distance (r’s range: -.23 to -.14; p’s >.10). Results suggest that self-reports of pain and disability may not correspond well with subjects’ actual capacities to perform simple everyday functions. Even the pain intensity increases evoked during the functional tests correlated weakly with performance, indicating that those reporting greatest pain increases during the functional tests did not differ on performance from those reporting small pain increases. Thus, having subjects perform simple functional tests that do not require complex apparatus could add greatly to the clinical picture portrayed by OA patients. Supported by a research grant from Forest Laboratories, Inc. (110) Withdrawn A04 Clinical Outcomes Measurement (111) Stanford-NIH Pain Registry: catalyzing the rate limited step of psychometrics withmodern patient-reported outcomes M Kao, K Cook, G Olson, T Pacht, B Darnall, S Weber, and S Mackey; Stanford School of Medicine, Palo Alto, CA Unlike passive biometric measurements, psychometric measures require active participation from subjects and are rate-limited by subject burden.We develop a comprehensive system of algorithms on the Stanford-NIH Pain Registry to support modern patient reported outcome (PRO) with item-response theory (IRT) and computerized adaptive testing (CAT). This system, called SNAPLCAT, is designed as amulti-feature computation engine for the NIH funded psychometric item banks NIH PROMIS and NIH Toolbox. Implemented on open sourceMEAN stackwith D3.js visualization, the system’s features include initialization (individualized or patient population priors), item selection (expected Kullback-Leibler, minimum expected posterior variance), advanced item selection (alpha-stratification, exposure control, content balancing, probabilistic constrained optimization), stopping rule (predicted standard error reduction, percentile width, hybrid), and estimation (expected a posteriori, maximum likelihood, maximum a posterior). Item banks and linkages are obtained from Northwestern Access Center and PROsetta Stone. Performance in 4,466 measurements in the Registry are analyzed. We find that basic CAT provided significant reduction in burden (mean number of items 6 SD, fold reduction): Anger (6.24+/-1.21, 4.6-fold vs BPAQ), Anxiety (4.93+/-0.97, 1.4-fold vs GAD-7), Depression (4.97+/-1.07, 1.8-fold vs PHQ-9), Fatigue (4.78+/-0.76, 8.4-fold vs FACIT-F), Physical Function (4.11+/-0.48, 4.9-fold vs HAQ-DI), Pain Interference (4.19+/-0.71, 1.7-fold vs BPI), Sleep Disturbance (4.95+/-1.41, 2.4-fold vs SDQ), Sleep-Related Impairment (4.54+/-1.24, 1.8-fold vs ESS). Altogether, the 132 classic instrument items may be alternatively assessed by 38.7 +/7.9 items, for 2.8 to 4.3 fold reduction in patient burden. In conclusion, using IRT and advanced CAT, the Stanford-NIH Pain Registry and SNAPL-CAT leverage the powers of NIH PROMIS and Toolbox, and enable big data psychometrics for the study of pain.