Performance Deterioration of Personalized Mood Prediction Using Passive Sensing

Zhou, Xinyi (2022) Performance Deterioration of Personalized Mood Prediction Using Passive Sensing. Bachelor thesis, Psychology.

	Text BachelorThesis_XZhou.pdf Restricted to Repository staff only Download (1MB)
	Text PermissionExamCommittee.pdf Restricted to Repository staff only Download (125kB)

Abstract

The time dynamics of affect are better captured by the Experience Sampling Method (ESM) through repeated measurements compared to traditional measurement approaches. However, the ESM can impose a substantial burden on the individual. More recently, studies have explored the use of passively collected data via smartphone or wearable sensors to predict fluctuations in mood, with the hope of replacing the ESM to reduce the burden. However, an important but thus far overlooked question is for how long a model is able to continuously and accurately predict mood before retaining is required. To address this, we utilized personalized models predicting positive and negative affect densely reported by participants (N = 10) based on passive smartphone data. We evaluated the predictions at varying temporal distances from the model training phase over the course of a day. On average across participants, the predictions achieved an R^2 of 0.182 (95% CI [0.088, 0.276]) for negative affect and 0.113 (95% CI [0.038, 0.188]) for positive affect at the first time point after the training phase. Importantly, deteriorations in prediction performance were observed over time for all participants. However, our findings may be limited by the small samples used for model training. Further investigations are warranted using larger training sizes and preferably models with improved prediction power. While more evidence is needed, our results suggested that regular active assessment and model updates could be expected for long-term monitoring of affect, due to the likely deterioration of prediction performance over time. We further argued that specific criteria regarding prediction performance for clinical use are required, before the clinical utility of the models can be determined.

Item Type:	Thesis (Bachelor)
Supervisor name:	Bringmann, L.F.
Degree programme:	Psychology
Differentiation route:	None [Bachelor Psychology]
Date Deposited:	29 Jul 2022 08:00
Last Modified:	29 Jul 2022 08:00
URI:	http://gmwpublic.studenttheses.ub.rug.nl/id/eprint/1203

Actions (login required)

View Item