Samuel Eke

Background/Objectives: Cancer survivors admitted to inpatient rehabilitation suffer from a complex burden of interrelated physical and psychological symptoms. While mean-level improvements during rehabilitation are well-documented, it remains unknown whether rehabilitation modifies the underlying structure of symptom interconnections—the symptom network—beyond reducing individual symptom scores. This study aimed to characterize symptom network structure at admission and discharge of a 21-day inpatient cancer rehabilitation program based on cancer-related physical symptoms and psychosocial functioning, formally compare network topology across timepoints, identify structurally central treatment targets, and assess the transdiagnostic generalizability of findings. Methods: Secondary analysis of routinely collected, electronic patient-reported outcome (PRO) data from 5066 cancer survivors (mean age 60.3 years, SD 12.2; 64.2% female; most frequent diagnoses: breast cancer = 36.9%, hematological malignancies = 10.4%; prostate cancer = 8.5%) admitted to a single-center inpatient rehabilitation program was performed between January 2017 and November 2022. The EORTC QLQ-C30 and the Hospital Anxiety and Depression Scale (HADS) questionnaires were utilized. Bayesian Gaussian Graphical Models were estimated at admission (T0) and discharge (T1) across 17 symptom and functioning domains using Bayesian Model Averaging (15,000 iterations). Edge-level change was quantified via posterior distributions of pairwise differences with 95% Highest Density Intervals. Node-level changes were assessed using Bayesian paired t-tests. Centrality was quantified by Expected Influence and Bridge Expected Influence. Results: Patients showed clinically meaningful improvements across all 17 domains during rehabilitation (all Bayes Factors >10; posterior probability of direction >99.9%). The largest standardized effects were observed for emotional functioning (Cohen’s d = 0.76), global health status (d = 0.69), and fatigue (d = 0.53). These improvements were clinically meaningful for a substantial proportion of patients: 62% improved by at least the minimal important difference in fatigue and 58% in emotional functioning, and the proportion of patients with probable anxiety fell from 15% to 6% and probable depression from 10% to 4%. Emotional functioning and anxiety were the most central domains in the symptom network—most strongly connected to the rest of patients’ symptom burden—at both admission and discharge. Despite the clinical improvements, the overall architecture of symptom interconnections changed little (83% of connections were unchanged). This indicates that the severity of symptoms was mitigated while the structure linking them together remained largely intact. The one connection that strengthened was that between impaired social functioning and financial difficulties (Δ = −0.112). Structural findings were consistent across ten cancer types (leave-one-out r > 0.80 in seven of ten). Conclusions: Over the course of inpatient cancer rehabilitation, patients showed large improvements against a background of largely stable symptom network architecture. Emotional functioning and anxiety occupy structurally central positions at both admission and discharge, identifying them as candidate domains warranting further investigation for network-informed rehabilitation. These findings provide a novel structural perspective on oncological rehabilitation and a framework for developing more targeted intervention strategies.

Childhood maltreatment (CM) is typically assessed using cumulative count scores that discard nuanced information about severity, frequency, duration, timing and experiential composition. This study evaluated whether similarity-based profile representations improve prediction of adult psychopathology and introduces a framework generalizing to multimodal data integration in translational psychiatry. A representative sample of 2,515 German adults completed measures of CM (ICAST-R) and psychopathology (ITQ; PHQ-4). Age-related severity trajectories were modeled using Bayesian regression. Cosine similarity-based profile representations were compared against conventional approaches using leave-one-out cross-validation with Bayesian bootstrap inference. Substantial within-category heterogeneity was evident across maltreatment experiences (e.g., emotional abuse severity: 0.40–0.68; physical neglect severity: 0.22–0.52). Sexual abuse showed consistent evidence for age-dependent severity, with earlier occurrences rated as more harmful (all P(β < 0) > 0.95). Profile-based representations outperformed frequency-based scoring by 53–91% and multiplicity-based scoring by 31–46% in cross-validated explained variance across outcomes (all P ≈ 1.00). The largest gains over severity-based models emerged for disturbances in self-organization (cvR² = 0.23; ΔcvR² = 0.05, + 25% over severity, P = 1.00), while core PTSD symptoms depended more on overall exposure intensity than experiential structure. Similarity-based representations substantially improved prediction of adult psychopathology, with the strongest gains for symptoms distinguishing complex from classic PTSD. The representationally agnostic framework extends naturally to multimodal designs integrating psychological, biological, and clinical data, addressing a key barrier in personalized medicine and precision psychiatry.