Inferring visual behaviour from user interaction data on a medical dashboard
1 like243 views
The study investigates the relationship between user interaction data and visual behavior on a medical dashboard to enhance decision-making in healthcare. It explores information overload and identifies adaptive user interface needs through analyzing interaction logs and eye-tracking data from laboratory and remote participants. The results suggest a correlation between interactive and visual behaviors, indicating potential for inferring visual behavior from user interaction data.
Inferring visual behaviour from user interaction data on a medical dashboard
- 1. Inferring Visual Behaviour from User Interaction Data on a Medical Dashboard! Ainhoa Yera, Javier Muguerza, Olatz Arberlaitz, Iñigo Perona Faculty of Informatics – University of the Basque Country UPV/EHU Richard Keers, Darren Ashcroft Division of Pharmacy and Optometry – University of Manchester Caroline Jay, Markel Vigo School of Computer Science – University of Manchester Richard Williams, Niels Peek Division of Informatics, Imaging and Data Sciences – University of Manchester Paper available at https://doi.org/10.1145/3194658.3194676
- 2. Motivation! • Health data + software • Proactive management of population health care • SMASH a primary care intervention – Medical dashboard – Patient safety – PINCER indicators1 • Conditions vs medications: CKD and NSAIDs • Demographics vs medications: Woman, smoker, +35, CHC – Pharmacists and GPs – Deployed in Salford (UK) 1: Avery et al. (2012) A pharmacist-lead information technology intervention for medication errors. The Lancet 379 1
- 3. SMASH screens! 5 - Trends 3 - Visualisation 1 - Landing page 4 - Patients 6 - Information 2 - Indicators 2
- 4. Problem! • Information overload problem • Addressing high information density in medical decision making tools? • “The right amount of information” • Adaptive user interfaces – What are these information needs? – What’s taxing decision making of clinicians? • Visual behaviour is proxy of interest and information overload “I just need the right amount of information” Louise, GP at Greater Manchester 3
- 5. Visual behaviour as a proxy of interest and information overload! 4
- 6. Problem! • Adaptive user interfaces – What are these information needs? – What’s taxing decision making of clinicians? • Visual behaviour is proxy of interest and information overload • Eye-trackers not expected beyond the lab setting • RQ: can we link user interaction data to visual behavior? “I just need the right amount of information” Louise, GP at Greater Manchester 5
- 7. Collected data! • Laboratory (N=6) – Interaction logs and eye-tracking data – Typical tasks • Remote (N=35) – Interaction logs – 10 months data (Jan-Oct 2016) 6
- 8. Computed metrics! • User interaction metrics – Exploration: mouse hovers between mouse clicks – Pace: elapsed time between two consecutive mouse clicks • Gaze metrics – Fixation duration: values between 50-400ms • Vectors of aggregations – V1: Exploration and pace on SMASH (length: 2) – V2: Exploration and pace per screen (length: 7x2) – G: Median fixation duration per AOI (length: 9) 7
- 9. Analysis: laboratory participants! • Finding individuals with similar interactive behaviours – Clustering analysis on log data (V1 and V2) • k=3 optimal when applying silhouette analysis on Cluster Validity Indexes, k: 3..10 • Euclidean distance • Same results for V1 and V2 • Hierarchical clustering including the centroids • Neighbour-joining method k-means, k=3, Euclidean Log data: clustering on V Log data: clustering on V P3 P1 P4 P5 P6 P2 P1 P2 P3 P4P5 P6 P6 P2 C3 P4 P5 C1 P3 P1 C2 0.00.20.40.60.81.01.21.4 Cluster Dendrogram DM_dmedEu2 Height 8
- 10. Analysis: laboratory participants! • Finding individuals with similar interactive behaviours – Correlation analysis on fixation durations (G) – Same pairs as those emerging from interactive behaviour Log data: clustering on V Log data: clustering on V P3 P1 P4 P5 P6 P2 Log data: clustering on V Gaze data: correlation on G P3 P1 P4 P5 P6 P2 9
- 11. Analysis: lab + remote participants! • Finding individuals with similar interactive behaviours – Lab (N=6) and remote (N=35) altogether – Clustering analysis on log data, Euclidean distance – k=3 optimal when applying silhouette analysis on Cluster Validity Indexes, k: 3..10 All= lab + remote (N=41) Log data: clustering on V P1 P2 P3 P4P5 P6 P6 10
- 12. Laboratory (N=6) All= lab + remote (N=41) Gaze data Log data Main outcome/hypothesis: those individuals belonging to a cluster in the remote group have similar visual behaviours to that of laboratory users Log data Hypothesis: Gaze data 11
- 13. Outcomes! • In terms of interactive behaviours clusters indicate – P2 and P6: slow pace and high exploration – P1 and P3: high pace and high exploration – P4 and P5: high pace and low exploration • In terms of visual behaviours (ie cognitive load) – P2 and P6: less load when looking at indicators – P1, P3, P4, P5: high cognitive load on indicators and data table 12
- 14. Conclusions! • There is a relationship between interactive behaviour and visual behaviour • We could infer visual behaviour by monitoring interaction data • Results not completely conclusive but promising • We have a higher certainty 13
- 15. Questions? Inferring Visual Behaviour from User Interaction Data on a Medical Dashboard! 15 markel.vigo@manchester.ac.uk @markelvigo