The two ends of the spectrum: comparing chronic schizophrenia and premorbid latent schizotypy by actigraphy

Abstract

Motor activity alterations are key symptoms of psychiatric disorders like schizophrenia. Actigraphy, a non-invasive monitoring method, shows promise in early identification. This study characterizes Positive Schizotypy Factor (PSF) and Chronic Schizophrenia (CS) groups using actigraphic data from two databases. At Hauke Land University Hospital, data from patients with chronic schizophrenia were collected; separately, at the University of Szeged, healthy university students were recruited and screened for PSF tendencies toward schizotypy. Several types of features are extracted from both datasets. Machine learning algorithms using different feature sets achieved nearly 90-95% for the CS group and 70-85% accuracy for the PSF. By applying model-explaining tools to the well-performing models, we could conclude the movement patterns and characteristics of the groups. Our study indicates that in the PSF liability phase of schizophrenia, actigraphic features related to sleep are most significant, but as the disease progresses, both sleep and daytime activity patterns are crucial. These variations might be influenced by medication effects in the CF group, reflecting the broader challenges in schizophrenia research, where the drug-free study of patients remains difficult. Further studies should explore these features in the prodromal and clinical High-Risk groups to refine our understanding of the development of the disorder.

Keywords: Actigraphy; Disease development; Machine learning; Mental disease.

Fig. 1

Remaining numbers of the days of each participant after data filtering. The Norwegian dataset is represented with colors of red (CS) and green (Control), and the Szeged dataset with orange (PSF) and blue (Control)

Fig. 2

Distribution of the detected sleep time, broken down by groups

Fig. 3

Slice of the activity data with sleep movement features. Horizontal line is the median/upper quartile boundary

Fig. 4

Illustration showing the wavelet-based analysis of nocturnal activity structures. a Activities of concatenated sleep periods for five consecutive nights. b Correlation-coefficient map of (a)’s time series, as determined by continuous wavelet analysis. c Structure parameters derived from the map in (b), as a function of the scale parameter for the 1–200 s time window. d Distribution of integrated structure factors (structure_pms) across the two volunteer groups (Control Group (blue) and Positive Schizotypy Factor Group (red))

Fig. 5

The CFFS steps are presented in a flowchart, which F represents the features. It begins by selecting features based on Welch’s test for statistical significance and mutual information, choosing only those that meet specific criteria for both. It then constructs a complete weighted graph using pairwise Pearson correlations among the selected features ($F_{e_1-e_m}$ is a feature subset of $F_{1-n}$ full feature set). Edges exceeding a certain correlation threshold are deleted to reduce inter-feature correlation. The process concludes by identifying cliques within this graph, which represent optimal combinations of less intercorrelated features. For example, $F_{i_1-i_4}$ is a 4-element whole subgraph (Clique) and will be a feature candidate; however, $F_{e_5}$ was left out from the analysis because it does not have enough low correlation features. Those Clique were used for training machine learning algorithms with a focus on maximizing accuracy and interpretability through Shapley values in a 3-fold cross-validation framework

Fig. 6

The proposed algorithm is represented by an operation graph, illustrating two sequential steps. It constructs candidate model configurations and selects the feature that results in the smallest estimation error or highest accuracy in the actual extension. The graph nodes are labeled with feature indices represented as numbers and enclosed in frames of different colors. The color variations signify the state of the examined variables. Specifically, features with black frames represent the already selected feature set in the current state, while features with colorful frames (red, green) are the candidate features in the set. The green-framed feature indicates the best feature, which exhibits the smallest estimation error or the highest accuracy compared to other potential variables enclosed in red frames. The directed edges in the graph represent transitions between different states of feature subsets. These subsets correspond to various predetermined feature selection methods employed in this context

Fig. 7

Results of the three algorithms on AHFS and CFSS. Within the Clique Forming Feature Selection (CFFS) framework, models were selected based on achieving at least 60% accuracy. In the USD, Logistic Regression saw 942 models exceed this threshold, while the HUHD had 559. For Random Forest, the numbers were 42 in the USD and 811 in the HUHD. Neural networks had 868 successful models in the USD and 634 in the HUHD. In contrast, the Adaptive Hybrid Feature Selection (AHFS) conducted 20 independent runs, each involving 20 iterative steps, ultimately identifying 400 features. The selection focused on pinpointing the most compact and high-performing feature sets, resulting in 20 optimal models per group

Fig. 8

Shapley values of the logistic regression (left: USD, right: HUHD)

Fig. 9

Shapley values of the Random Forest (left: USD, right: HUHD)

Fig. 10

Shapley values of the ANN (left: USD, right: HUHD)

Fig. 11

Shapley values of the ANN(AHFS) (left: USD, right: HUHD)