Interpretable machine learning for psychological research: Opportunities and pitfalls.

In recent years, machine learning methods have become increasingly popular prediction methods in psychology. At the same time, psychological researchers are typically not only interested in making predictions about the dependent variable, but also in learning which predictor variables are relevant, how they influence the dependent variable, and which predictors interact with each other. However, most machine learning methods are not directly interpretable. Interpretation techniques that support researchers in describing how the machine learning technique came to its prediction may be a means to this end. We present a variety of interpretation techniques and illustrate the opportunities they provide for interpreting the results of two widely used black box machine learning methods that serve as our examples: random forests and neural networks. At the same time, we illustrate potential pitfalls and risks of misinterpretation that may occur in certain data settings. We show in which way correlated predictors impact interpretations with regard to the relevance or shape of predictor effects and in which situations interaction effects may or may not be detected. We use simulated didactic examples throughout the article, as well as an empirical data set for illustrating an approach to objectify the interpretation of visualizations. We conclude that, when critically reflected, interpretable machine learning techniques may provide useful tools when describing complex psychological relationships. (PsycInfo Database Record (c) 2023 APA, all rights reserved).

Successful weaning versus permanent cerebrospinal fluid diversion after aneurysmal subarachnoid hemorrhage: post hoc analysis of a Swiss multicenter study.

OBJECTIVE: Acute hydrocephalus is a frequent complication after aneurysmal subarachnoid hemorrhage (aSAH). Among patients needing CSF diversion, some cannot be weaned. Little is known about the comparative neurological, neuropsychological, and health-related quality-of-life (HRQOL) outcomes in patients with successful and unsuccessful CSF weaning. The authors aimed to assess outcomes of patients by comparing those with successful and unsuccessful CSF weaning; the latter was defined as occurring in patients with permanent CSF diversion at 3 months post-aSAH. METHODS: The authors included prospectively recruited alert (i.e., Glasgow Coma Scale score 13-15) patients with aSAH in this retrospective study from six Swiss neurovascular centers. Patients underwent serial neurological (National Institutes of Health Stroke Scale), neuropsychological (Montreal Cognitive Assessment), disability (modified Rankin Scale), and HRQOL (EuroQol-5D) examinations at < 72 hours, 14-28 days, and 3 months post-aSAH. RESULTS: Of 126 included patients, 54 (42.9%) developed acute hydrocephalus needing CSF diversion, of whom 37 (68.5%) could be successfully weaned and 17 (31.5%) required permanent CSF diversion. Patients with unsuccessful weaning were older (64.5 vs 50.8 years, p = 0.003) and had a higher rate of intraventricular hemorrhage (52.9% vs 24.3%, p = 0.04). Patients who succeed in restoration of physiological CSF dynamics improve on average by 2 points on the Montreal Cognitive Assessment between 48-72 hours and 14-28 days, whereas those in whom weaning fails worsen by 4 points (adjusted coefficient 6.80, 95% CI 1.57-12.04, p = 0.01). They show better neuropsychological recovery between 48-72 hours and 3 months, compared to patients in whom weaning fails (adjusted coefficient 7.60, 95% CI 3.09-12.11, p = 0.02). Patients who receive permanent CSF diversion (ventriculoperitoneal shunt) show significant neuropsychological improvement thereafter, catching up the delay in neuropsychological improvement between 14-28 days and 3 months post-aSAH. Neurological, disability, and HRQOL outcomes at 3 months were similar. CONCLUSIONS: These results show a temporary but clinically meaningful cognitive benefit in the first weeks after aSAH in successfully weaned patients. The resolution of this difference over time may be due to the positive effects of permanent CSF diversion and underlines its importance. Patients who do not show progressive neuropsychological improvement after weaning should be considered for repeat CT imaging to rule out chronic (untreated) hydrocephalus.

Trajectories of reproductive transition phase mood disorder from pregnancy to postpartum: A Swiss longitudinal study.

BACKGROUND: Depressive symptoms are common in the peripartum period and pose a great risk to the well-being of the mother, the infant, and the entire family. Evidence from longitudinal studies suggests that affected women do not constitute one homogeneous group in terms of severity, chronicity, and onset of symptoms. To account for individual differences regarding the longitudinal course of depressive symptoms from pregnancy to the postpartum period, growth mixture models have proven to be useful. METHODS: We conducted a group-based trajectory modeling analysis to identify perinatal depressive symptom trajectories in a Swiss sample (n = 151). Depressive symptoms were assessed six times, covering nearly 6 months from the third trimester of pregnancy to 3 months postpartum. In addition to determining perinatal depressive symptom trajectories, we aimed to examine whether these trajectories are linked to psychopathological risk factors such as a history of premenstrual syndrome (PMS), anxiety, prenatal stress, and somatic symptoms after delivery that are associated with hormonal fluctuations. RESULTS: The findings revealed three trajectories of perinatal depressive symptoms that were relatively stable over time and differed in symptom load (low, medium, high), as well as one trajectory of decreasing symptoms, with a significant symptom reduction after giving birth. Women with a higher depressive symptom load experienced a greater degree of prior premenstrual symptoms, prenatal anxiety, and birth anxiety, as well as somatic symptoms after delivery. CONCLUSION: Further research is needed to account for the distinct trajectories of perinatal depressive symptoms in order to provide appropriate care for affected women. A focus on somatic symptoms after delivery and their association with depressive mood is essential to better understand the potential shared etiopathology of reproductive transition phase mood disorders.

Longitudinal neuropsychological assessment after aneurysmal subarachnoid hemorrhage and its relationship with delayed cerebral ischemia: a prospective Swiss multicenter study.

OBJECTIVE: While prior retrospective studies have suggested that delayed cerebral ischemia (DCI) is a predictor of neuropsychological deficits after aneurysmal subarachnoid hemorrhage (aSAH), all studies to date have shown a high risk of bias. This study was designed to determine the impact of DCI on the longitudinal neuropsychological outcome after aSAH, and importantly, it includes a baseline examination after aSAH but before DCI onset to reduce the risk of bias. METHODS: In a prospective, multicenter study (8 Swiss centers), 112 consecutive alert patients underwent serial neuropsychological assessments (Montreal Cognitive Assessment [MoCA]) before and after the DCI period (first assessment, < 72 hours after aSAH; second, 14 days after aSAH; third, 3 months after aSAH). The authors compared standardized MoCA scores and determined the likelihood for a clinically meaningful decline of ≥ 2 points from baseline in patients with DCI versus those without. RESULTS: The authors screened 519 patients, enrolled 128, and obtained complete data in 112 (87.5%; mean [± SD] age 53.9 ± 13.9 years; 66.1% female; 73% World Federation of Neurosurgical Societies [WFNS] grade I, 17% WFNS grade II, 10% WFNS grades III-V), of whom 30 (26.8%) developed DCI. MoCA z-scores were worse in the DCI group at baseline (-2.6 vs -1.4, p = 0.013) and 14 days (-3.4 vs -0.9, p < 0.001), and 3 months (-0.8 vs 0.0, p = 0.037) after aSAH. Patients with DCI were more likely to experience a decline of ≥ 2 points in MoCA score at 14 days after aSAH (adjusted OR [aOR] 3.02, 95% CI 1.07-8.54; p = 0.037), but the likelihood was similar to that in patients without DCI at 3 months after aSAH (aOR 1.58, 95% CI 0.28-8.89; p = 0.606). CONCLUSIONS: Aneurysmal SAH patients experiencing DCI have worse neuropsychological function before and until 3 months after the DCI period. DCI itself is responsible for a temporary and clinically meaningful decline in neuropsychological function, but its effect on the MoCA score could not be measured at the time of the 3-month follow-up in patients with low-grade aSAH with little or no impairment of consciousness. Whether these findings can be extrapolated to patients with high-grade aSAH remains unclear. Clinical trial registration no.: NCT03032471 (ClinicalTrials.gov).

Dissociation of motor control from motor awareness in awake sleepwalkers: An EEG study in virtual reality.

Recent behavioral evidence from a virtual reality (VR) study indicates that awake sleepwalkers show dissociation of motor control and motor awareness. This dissociation resembles the nocturnal disintegration of motor awareness and movement during episodes of sleepwalking. Here, we set out to examine the neural underpinnings of altered motor awareness in sleepwalkers by measuring EEG modulation during redirected walking in VR. To this end, we measured scalp EEG during ongoing motor behavior to provide information on motor processing and its modulation in VR. Using this approach, we discovered distinct EEG patterns associated to dual tasking and sub-threshold motor control in sleepwalkers compared to control subjects. These observations provide further electrophysiological evidence for the proposed brain-body dissociation in awake sleepwalkers. This study shows proof-of-principle that EEG biomarkers of movement in a VR setting add to the understanding of altered motor awareness in sleepwalkers. In a broader perspective, we confirm the feasibility of using the additional dimensionality in VR providing novel diagnostic biomarkers not accessible to conventional clinical investigations. In future studies, this approach could contribute to the diagnostic work-up of patients with a broad spectrum of neurological diseases.

Walking through virtual mazes: Spontaneous alternation behaviour in human adults.

Spontaneous alternation behaviour (SAB) is the tendency to systematically alternate directional choices in successive maze arms. Originally discovered in rats, SAB has been extensively investigated in a broad range of species. In humans, however, SAB has been mostly ignored, possibly due to the difficulties arising from the use of life-size mazes. We here propose to close this gap by advancing the study of human SAB by use of virtual reality (VR). Alternation rates in humans were examined in three experiments, each deploying a specific type of virtual maze. The three virtual mazes tested 1) the effect of a concurrent cognitive task on baseline alternation rates, 2) the differential influence of locomotor and visual factors on alternation behaviour, and 3) the direction alternation in an unrestricted open space. We report a general tendency in adult human walkers to alternate walking directions in the classical T-maze context. The search for an effect of a concurrent cognitive task and the influence of locomotor and visual factors on alternation behaviour remained inconclusive. No evidence for alternation behaviour in an open space was found. Together, the experimental series elucidates the presence and characteristics of SAB in humans and paves the way for the systematic study of its neurocognitive basis.

Influence of the Intensive Care Unit Environment on the Reliability of the Montreal Cognitive Assessment.

Background: Neuropsychological screening becomes increasingly important for the evaluation of subarachnoid hemorrhage (SAH) and stroke patients. It is often performed during the surveillance period on the intensive (ICU), while it remains unknown, whether the distraction in this environment influences the results. We aimed to study the reliability of the Montreal Cognitive Assessment (MoCA) in the ICU environment. Methods: Consecutive stable patients with recent brain injury (tumor, trauma, stroke, etc.) were evaluated twice within 36 h using official parallel versions of the MoCA (ΔMoCA). The sequence of assessment was randomized into (a) busy ICU first or (b) quiet office first with subsequent crossover. For repeated MoCA, we determined sequence, period, location effects, and the intraclass correlation coefficient (ICC). Results: N = 50 patients were studied [n = 30 (60%) male], with a mean age of 57 years. The assessment's sequence ["ICU first" mean ΔMoCA -1.14 (SD 2.34) vs. "Office first" -0.73 (SD 1.52)] did not influence the MoCA (p = 0.47). On the 2nd period, participants scored 0.96 points worse (SD 2.01; p = 0.001), indicating no MoCA learning effect but a possible difference in parallel versions. There was no location effect (p = 0.31) with ΔMoCA between locations (Office minus ICU) of -0.32 (SD 2.21). The ICC for repeated MoCA was 0.87 (95% CI 0.79-0.92). Conclusions: The reliability of the MoCA was excellent, independent from the testing environment being ICU or office. This finding is helpful for patient care and studies investigating the effect of a therapeutic intervention on the neuropsychological outcome after SAH, stroke or traumatic brain injury.

Visual capture of gait during redirected walking.

Redirected walking allows users of virtual reality applications to explore virtual environments larger than the available physical space. This is achieved by manipulating users' walking trajectories through visual rotation of the virtual surroundings, without users noticing this manipulation. Apart from its applied relevance, redirected walking is an attractive paradigm to investigate human perception and locomotion. An important yet unsolved question concerns individual differences in the ability to detect redirection. Addressing this question, we administered several perceptual-cognitive tasks to healthy participants, whose thresholds of detecting redirection in a virtual environment were also determined. We report relations between individual thresholds and measures of multisensory weighting (visually-assisted postural stability (Romberg quotient), subjective visual vertical (rod-and-frame test) and illusory self-motion (vection)). The performance in the rod-and-frame test, a classical measure of visual dependency regarding postural information, showed the strongest relation to redirection detection thresholds: The higher the visual dependency, the higher the detection threshold. This supports the interpretation of users' neglect of redirection manipulations as a "visual capture of gait". We discuss how future interdisciplinary studies, merging the fields of virtual reality and psychology, may help improving virtual reality applications and simultaneously deepen our understanding of how humans process multisensory conflicts during locomotion.