Fig 1.
Three sample patients’ CCFs between time series.
The first row is the records of the two series (PHQ-8 and Item 9) for each patient and their fitted curve; the second row is the CCFs between PHQ-8 and Item 9. One unit of time is two weeks.
Fig 2.
Histogram of the lag k at maximum CCF in the population, PHQ-8 vs Item 9.
We excluded 214 patients with zero CCF between PHQ-8 and Item 9 for all lags between -5 and 5.
Fig 3.
The association between the changes of PHQ-8 and Item 9 in both short term (less than 1 month) and long term (at 3, 6, 9 months).
The solid lines are the result of linear regression.
Table 1.
Spearman’s rank-order correlation and linear regression for PHQ-8 and Item 9.
Table 2.
Subgroup size for patients with various PHQ-8 and Item 9 changing patterns.
Pattern (a): PHQ-8 increases and Item 9 decreases; Pattern (b): PHQ-8 decreases and Item 9 increases.
Fig 4.
The subtype analysis result with hidden structures learned from the PHQ-8 trajectories.
(a) Latent patterns learned from the PHQ-8 data. These patterns are visualized as the rows Wi. In each panel, the x-axis is the time with a period of 2 weeks, and the y-axis represents the PHQ-8 scores of each basis trajectory. (b) Embed the activation h (25 dimensions) of each patient into 2-dimensional space with t-SNE and cluster them with the k-means algorithm in the 2-dimensional space. (c) The value of activation h on each latent pattern (25 columns) of each patient (610 rows) after reordering the rows by the clustering. (d) Mean trajectories of average PHQ-8 and Item 9 by groups, using the clustering results. One unit of time is two weeks. We used the average score of the first 8 questions in the PHQ to represent PHQ-8, which has the same range of 0 to 3 to Item 9.