Is Fidgety Philip's ground truth also ours? The creation and application of a machine learning algorithm. (December 2020)

• Record Type: Journal Article
• Authors: Beyzaei, Nadia; Bao, Seraph; Bu, Yanyun; Hung, Linus; Hussaina, Hebah; Maher, Khaola Safia; Chan, Melvin; Garn, Heinrich; Kloesch, Gerhard; Kohn, Bernhard; Kuzeljevic, Boris; McWilliams, Scout; Spruyt, Karen; Tse, Emmanuel; Machiel Van der Loos, Hendrik F.; Kuo, Calvin; Ipsiroglu, Osman S.
• Is Part Of: Journal of psychiatric research. Volume 131(2020)
• Page Start: 144
• Page End: 151
• Publication Date: 2020-12
• Subjects: Movement disorders -- Sleep-related movement disorders -- Misdiagnosis -- Over-medication -- Adverse drug reactions; Psychiatry -- Periodicals; Mental Disorders -- Periodicals; Maladies mentales -- Périodiques; Psychiatry; Electronic journals; Periodicals; 616.89005
• DOI: 10.1016/j.jpsychires.2020.08.033
• Languages: English
• ISSNs: 0022-3956

Abstract

Abstract: Background: Behavioral observations support clinical in-depth phenotyping but phenotyping and pattern recognition are affected by training background. As Attention Deficit Hyperactivity Disorder, Restless Legs syndrome/Willis Ekbom disease and medication induced activation syndromes (including increased irritability and/or akathisia), present with hyperactive-behaviors with hyper-arousability and/or hypermotor-restlessness (H-behaviors), we first developed a non-interpretative, neutral pictogram-guided phenotyping language (PG-PL) for describing body-segment movements during sitting. Methodology & results: The PG-PL was applied for annotating 12 1-min sitting-videos (inter-observer agreements >85%->97%) and these manual annotations were used as a ground truth to develop an automated algorithm using OpenPose, which locates skeletal landmarks in 2D video. We evaluated the algorithm's performance against the ground truth by computing the area under the receiver operator curve (>0.79 for the legs, arms, and feet, but 0.65 for the head). While our pixel displacement algorithm performed well for the legs, arms, and feet, it predicted head motion less well, indicating the need for further investigations. Conclusion: This first automated analysis algorithm allows to start the discussion about distinct phenotypical characteristics of H-behaviors during structured behavioral observations and may support differential diagnostic considerations via in-depth phenotyping of sitting behaviors and, in consequence, of better treatment concepts.