Nevin Manimala

PLoS One. 2025 Oct 9;20(10):e0333379. doi: 10.1371/journal.pone.0333379. eCollection 2025.

ABSTRACT

We propose TIC-FusionNet, a trend-aware multimodal deep learning framework for time series forecasting with integrated visual signal analysis, aimed at addressing the limitations of unimodal and short-range dependency models in noisy financial environments. The architecture combines Exponential Moving Average (EMA) decomposition for denoising and trend extraction, a lightweight Linear Transformer for efficient long-sequence temporal modeling, and a spatial-channel CNN with CBAM attention to capture morphological patterns from candlestick chart images. A gated fusion mechanism adaptively integrates numerical and visual modalities based on context relevance, enabling dynamic feature weighting under varying market conditions. We evaluate TIC-FusionNet on six real-world stock datasets, including four major Chinese and U.S. companies-Amazon, Tesla, Kweichow Moutai, Ping An Insurance, China Vanke-and Apple-covering diverse market sectors and volatility patterns. The model is compared against a broad range of baselines, including statistical models (ARIMA), classical machine learning methods (Random Forest, SVR), recurrent and convolutional neural networks (LSTM, TCN, CNN-only), and recent Transformer-based architectures (Informer, Autoformer, Crossformer, iTransformer). Experimental results demonstrate that TIC-FusionNet achieves consistently superior predictive accuracy and generalization, outperforming state-of-the-art baselines across all datasets. Extensive ablation studies verify the critical role of each architectural component, while attention-based interpretability analysis highlights the dominant technical indicators under different volatility regimes. These findings not only confirm the effectiveness of multimodal integration in capturing complementary temporal-visual cues, but also provide valuable insights into model decision-making. The proposed framework offers a robust, scalable, and interpretable solution for multimodal temporal prediction tasks, with strong potential for deployment in intelligent forecasting, sensor fusion, and risk-aware decision-making systems.

PMID:41066756 | DOI:10.1371/journal.pone.0333379

JMIR Res Protoc. 2025 Oct 9;14:e74295. doi: 10.2196/74295.

ABSTRACT

BACKGROUND: Aggression and violence are prevalent in forensic psychiatric inpatient care. These behaviors significantly impact treatment outcomes, create challenging work environments for staff, and strain relationships between patients and caregivers. Managing such behaviors poses a formidable challenge that necessitates innovative approaches and evidence-based interventions. The Therapeutic Meeting with Aggression (TERMA) model is a staff training program designed to equip staff with strategies to de-escalate patient aggression, thus reducing violence and increasing patients’ and staffs’ perceived safety.

OBJECTIVE: The aim of this project is to evaluate the violence prevention model TERMA regarding perceived safety by patients and staff and adverse events within forensic psychiatric inpatient care. In addition, the project will investigate whether the organizational culture affects the implementation of the TERMA model.

METHODS: The project includes an observational study with a before and after design. Implementation of the TERMA model consists of an 8-seminar staff training program. Data sources include questionnaires, medical records, and registries. Quantitative data will be analyzed using descriptive and comparative statistics. To analyze changes between measurements, dependent sample 2-tailed t tests will be used for normally distributed data, and the Wilcoxon signed-rank test will be applied when normality is not met. The project will also include qualitative interview studies, which are planned to be analyzed using qualitative inductive content analysis.

RESULTS: Participant enrollment began in July 2023 and was concluded by the end of 2024. Data collection and analysis of quantitative data are expected to be completed by early 2026, after which the study findings will be submitted for publication in peer-reviewed scientific journals. Collection of qualitative data is scheduled for the second half of 2025 and 2026.

CONCLUSIONS: This study can add valuable knowledge about the effects of the violence prevention model TERMA.

PMID:41066754 | DOI:10.2196/74295

PLoS Biol. 2025 Oct 9;23(10):e3003431. doi: 10.1371/journal.pbio.3003431. Online ahead of print.

ABSTRACT

Decomposing behavior into elementary components remains a central challenge in computational neuroethology. The current standard in laboratory animals involves multi-view video tracking, which, while providing unparalleled access to full-body kinematics, imposes environmental constraints, is data-intensive, and has limited scalability. We present an alternative approach using inertial sensors, which capture high-resolution, environment-independent, compact 3D kinematic data, and are commonly integrated into rodent neurophysiological devices. Our analysis pipeline leverages unsupervised, computationally efficient change-point detection to break down inertial time series into variable-length, statistically homogeneous segments. These segments are then grouped into candidate behavioral motifs through high-dimensional, model-based probabilistic clustering. We demonstrate that this approach achieves detailed rodent behavioral mapping using head inertial data. Identified motifs, corroborated by video recordings, include orienting movements, grooming components, locomotion, and olfactory exploration. Higher-order behavioral structures can be accessed by applying a categorical hidden Markov model to the motif sequence. Additionally, our pipeline detects both overt and subtle motor changes in a mouse model of Parkinson’s disease and levodopa-induced dyskinesia, highlighting its utility for behavioral phenotyping. This methodology offers the possibility of conducting high-resolution, observer-unbiased behavioral analysis at minimal computational cost from easily scalable and environmentally unconstrained recordings.

PMID:41066739 | DOI:10.1371/journal.pbio.3003431

JCO Precis Oncol. 2025 Oct;9:e2500489. doi: 10.1200/PO-25-00489. Epub 2025 Oct 9.

ABSTRACT

PURPOSE: Lung cancer remains the leading cause of cancer-related mortality worldwide, with a substantial risk of recurrence even in early-stage non-small cell lung cancer (NSCLC) after curative surgery. Circulating tumor DNA (ctDNA)-based detection of minimal residual disease (MRD) has emerged as a promising tool for identifying patients at increased risk of relapse. However, the predictive effectiveness of ctDNA remains uncertain because of variability in study designs, detection strategies, and statistical power.

MATERIALS AND METHODS: We conducted a systematic meta-analysis of 30 studies involving 3,287 patients with postoperative NSCLC to evaluate the diagnostic performance of ctDNA-based MRD testing for recurrence detection and survival prediction. Eligible studies were identified through a comprehensive literature search and quality-assessed using the QUADAS-2 tool. Pooled diagnostic estimates were calculated using bivariate random-effects models. Subgroup analyses compared tumor-informed and tumor-agnostic detection strategies at both landmark and longitudinal postoperative time points.

RESULTS: In landmark analyses, tumor-informed assays demonstrated higher specificity (0.97 v 0.93) and AUC (0.81 v 0.70) than tumor-agnostic approaches, which showed slightly higher sensitivity (0.44 v 0.42). In longitudinal monitoring, the differences narrowed: tumor-informed assays retained higher specificity (0.96 v 0.88), whereas tumor-agnostic methods exhibited modestly higher sensitivity (0.79 v 0.76) and AUC (0.91 v 0.86).

CONCLUSION: Our findings indicate that ctDNA-based MRD testing provides clinically meaningful prognostic information for postoperative recurrence in early-stage NSCLC. Both detection strategies offer complementary strengths, with tumor-informed assays excelling in specificity and tumor-agnostic approaches offering greater sensitivity in some settings. These results highlight the potential of ctDNA MRD testing to enhance postoperative surveillance and guide personalized disease management in early-stage NSCLC.

PMID:41066727 | DOI:10.1200/PO-25-00489

Neurology. 2025 Nov 11;105(9):e214197. doi: 10.1212/WNL.0000000000214197. Epub 2025 Oct 9.

ABSTRACT

BACKGROUND AND OBJECTIVES: This analysis used clinical data from prospectively followed participants meeting criteria for probable dementia with Lewy bodies (DLB) in the Mayo Clinic Alzheimer’s Disease Research Center (ADRC) between 1998 and 2024. DLB is characterized by unique core features of visual hallucinations (VHs), parkinsonism, REM sleep behavior disorder, and cognitive fluctuations with a variable disease course. DLB is associated with a poor prognosis, but whether these unique DLB core clinical features influence survival is unknown. We aimed to determine whether core clinical features are associated with survival in patients with probable DLB.

METHODS: Patients followed in the Mayo Clinic ADRC between 1998 and 2024 underwent annual clinical assessments. Those who met clinical criteria for probable DLB were analyzed. Time-dependent Cox proportional hazard models using age as the time scale determined associations between the individual and cumulative number of core clinical DLB features and survival. The prognostic significance of core features present at the time DLB criteria were met was assessed in separate models. Models were adjusted for sex and duration from the onset of cognitive symptoms to DLB diagnosis.

RESULTS: Of 488 patients with probable DLB meeting inclusion criteria, 118 (24%) were women with a mean age of 71.9 ± 8.4 years at the time of meeting probable DLB criteria. Shorter survival was associated with the development of VHs (hazard ratio [HR] 3.25, 95% CI 2.46-4.29) and parkinsonism (HR 2.28, 95% CI 1.54-3.39) during the disease course and VHs at the time of DLB diagnosis (HR 1.60, 95% CI 1.18-2.16). All four core features were also associated with shorter survival (4 core features vs 2 core features, HR 3.58 95% CI 2.66-4.80, 4 core features vs 3 core features, HR 2.46, 95% CI 1.86-3.25). In 191 patients (45 women (24%) with a mean age of 71.2 ± 8.6 years at probable DLB diagnosis) with autopsy-confirmed DLB, VHs, parkinsonism, and all four core features were associated with shorter survival. Sex was not associated with survival.

DISCUSSION: VHs, parkinsonism, and the development of all 4 core features were associated with shorter survival in probable and in autopsy-confirmed DLB. These findings have important prognostic and management implications for patients with DLB and their caregivers.

PMID:41066723 | DOI:10.1212/WNL.0000000000214197

J Clin Oncol. 2025 Oct 9:101200JCO2501114. doi: 10.1200/JCO-25-01114. Online ahead of print.

ABSTRACT

BACKGROUND: Recent studies reveal that 5-18% of children with cancer harbor pathogenic variants in known cancer predisposing genes. However, DNA damage repair (DDR) genes, which are frequently somatically altered in pediatric tumors, have not been systematically examined as a source of novel cancer predisposing signals.

METHODS: To address this gap, we interrogated 189 DDR genes for presence of germline predisposing variants (PV) among 5,993 childhood cancer cases and 14,477 adult non-cancer controls (discovery cohort). PV were determined using a tiered approach incorporating ClinVar annotations, InterVar classification, and in silico tools (REVEL, CADD, MetaSVM). Using logistic and firth regression, we identified genes with PV statistically enriched in tumors and replicated findings among 1,497additional childhood cancer cases across three independent cohorts.

FINDINGS: Analysis across all cancers revealed enrichment of TP53 PV. Cancer-specific analyses confirmed known associations including germline TP53 PV in adrenocortical carcinoma, high-grade glioma (HGG), and medulloblastoma (MB), PMS2 in HGG and non-Hodgkin lymphoma (NHL), MLH1 in HGG, BRCA2 in NHL, and BARD1 in neuroblastoma. In addition, four novel associations were uncovered, including BRCA1 in ependymoma, SPIDR in HGG, SMC5 in MB, and SMARCAL1 in osteosarcoma (OS). Importantly, the SMARCAL1:OS association was significant in the discovery (6/230, 2.6%, FDR_logistic=0.0189) as well as all three replication cohorts (Childhood Cancer Survivor Study: 8/275, 2.9%; P_Fisher<0.0001; German Childhood Cancer Registry: 4/135, 3%, P_Fisher=0.002; INdividualized Therapy FOr Relapsed Malignancies in Childhood: 4/217, 1.8%, P_Fisher=0.012). The remaining wildtype SMARCAL1 allele was deleted in three of four OS tumors with available data.

INTERPRETATION: Our study confirms the relevance DDR genetic variation in pediatric cancer risk and establishes SMARCAL1 as a novel OS predisposing gene, providing insights into tumor biology and creating opportunities to optimize care for patients with this challenging tumor.

PMID:41066719 | DOI:10.1200/JCO-25-01114

Brief Bioinform. 2025 Aug 31;26(5):bbaf497. doi: 10.1093/bib/bbaf497.

ABSTRACT

Spatial transcriptomics (ST) technology enables the simultaneous capture of gene expression profile and spatial information within 2D tissue slices. However, conventional analyses that process each individual slice independently often overlook shared features across multiple slices, limiting comprehensive biological insights. To address this, we introduce GRASS, a deep graph representation learning-based framework designed for the integration and alignment of multislice ST data. GRASS consists of two core modules: GRASS_Integration, which employs a heterogeneous graph architecture integrating contrastive learning and a multi-expert collaboration strategy to fully utilize both shared and unique information, enabling multislice integration, clustering, and various downstream analyses; and GRASS_Alignment, which uses a dual-perception similarity metric to guide spot-level alignment, supporting downstream tasks such as imputation and 3D reconstruction. Experimental results on seven ST datasets from five different platforms demonstrate that GRASS consistently outperforms eight state-of-the-art methods in both integration and alignment tasks. By comprehensively addressing multi-level information integration, GRASS emerges as an ideal solution for the joint analysis of multislice ST data.

PMID:41066694 | DOI:10.1093/bib/bbaf497

JMIR Serious Games. 2025 Oct 9;13:e66673. doi: 10.2196/66673.

ABSTRACT

BACKGROUND: Cognitive training is an effective approach to support cognitive function in older adults. Incorporating meaningful leisure activities, such as gardening, may enhance both engagement and training outcomes. While fully immersive virtual reality (VR) offers ecologically valid and engaging environments that can further boost motivation, limited research has explored the combination of VR-based cognitive training and leisure activities for older adults.

OBJECTIVE: This study aims to assess the feasibility, usability, and preliminary effectiveness of leisure-based VR cognitive training for community-dwelling older adults.

METHODS: A fully immersive VR cognitive training system, controlled via a head-mounted display, was developed, incorporating gardening-themed activities such as planting, fertilizing, watering, and harvesting. These tasks were designed to engage multiple cognitive domains, including memory, attention, executive function, processing speed, and visuospatial abilities. The program consisted of 16 sessions delivered over 8 weeks (twice weekly, 1 hour per session). Cognitive outcomes were assessed before and after training using the Montreal Cognitive Assessment, the digit symbol substitution test, word list immediate and delayed recall, spatial span, and the Stroop Color and Word Test. Feasibility, acceptance, and usability were evaluated using the System Usability Scale and a posttraining questionnaire. Licensed occupational therapists from both community and institutional settings assessed the training system’s usability.

RESULTS: All 41 participants (mean age 69.79, SD 5.05 y) completed the training with 100% adherence and no serious adverse events. Feasibility ratings-particularly for perceived usefulness, intention to use, and subjective norms-reflected strong acceptance. Usability ratings from older adults indicated high ease of use, enjoyment, and positive experience, while professionals rated the system as moderately usable (mean System Usability Scale score 68.01, SD 8.38). Statistically significant improvements were observed in general cognition (P=.004), processing speed (P=.049), immediate and delayed memory (P<.001), and executive function (P=.002). No significant changes were found in visuospatial memory (P=.29).

CONCLUSIONS: This study provides preliminary evidence supporting the feasibility and usability of a gardening-based VR cognitive training program for older adults. Feasibility was demonstrated through full adherence, absence of major adverse events, and high participant acceptance. Usability feedback was favorable from both older adults and professionals across community and long term care settings. Additionally, improvements in multiple cognitive domains, including general cognition, processing speed, memory, and executive function, suggest potential cognitive benefits. Future randomized controlled trials with more diverse samples and extended follow-up are warranted to confirm and expand upon these findings.

PMID:41066691 | DOI:10.2196/66673

Age Ageing. 2025 Aug 29;54(10):afaf285. doi: 10.1093/ageing/afaf285.

ABSTRACT

OBJECTIVE: Older adults face high risk of falls. We developed an electronic-health-record (EHR) based machine-learning (ML) model to predict 1-year risk of fall in older adults for pre-emptive intervention.

METHODS: We included 4 902 161 records from 1 142 000 adults aged ≥65 years who attended the Hong Kong Hospital Authority (HA) facilities in 2013-2017. We included 260 predictors including demographics, in-patient/out-patient admissions, emergency department (ED) attendance, complications, medications and laboratory tests during 1-year period to predict fall events based on diagnostic codes in the ensuing 12 months. The cohort was randomly split into training, testing and internal validation sets in a 7:2:1 ratio. We evaluated the performance of six ML-algorithms.

RESULTS: 67 163 fall events were accrued with the XGBoost model having the best performance in the validation set (area-under-the-receiver-operating-characteristic-curve [AUROC] = 0.979, area-under-the-precision-recall-curve [AUPRC] = 0.764; positive-predictive-value [PPV] = 0.614) versus logistic-regression model (AUROC = 0.885, AUPRC = 0.169; PPV = 0.210). The top 30 predictors included number of ED attendance, fasting plasma glucose, number and types of outpatient appointments, ED triage category of ‘urgent’, number of admissions and stay, age, residential districts, history of fall and medication use with an AUROC of 0.939 in a validation cohort of patients with diabetes. In an age- and sex-matched sub-cohort, compared to the widely-used Morse Fall Score, XGBoost model had higher sensitivity (0.569-versus-0.139) with optimal balance of identifying positive cases whilst simultaneously minimising false positives and false negative (F1 score: 0.626-versus-0.555).

CONCLUSIONS: Our ML-model highlights the utility of EHR in identifying high-risk individuals for falls, supporting integrating into the EHR system for targeted preventive actions.

PMID:41066674 | DOI:10.1093/ageing/afaf285

Epilepsia. 2025 Oct 9. doi: 10.1111/epi.18667. Online ahead of print.

ABSTRACT

OBJECTIVE: This study was undertaken to develop a framework for focal cortical dysplasia (FCD) detection using surface-based morphometric (SBM) analysis and machine learning (ML) applied to three-dimensional (3D) magnetic resonance fingerprinting (MRF).

METHODS: We included 114 subjects (44 patients with medically intractable focal epilepsy and FCD, 70 healthy controls [HCs]). All subjects underwent high-resolution 3-T MRF scans generating T1 and T2 maps. All patients had clinical T1-weighted (T1w) images; 35 also had 3D fluid-attenuated inversion recovery (FLAIR). A 3D region of interest (ROI) was manually created for each lesion. All maps/images and lesion ROIs were registered to T1w images. Surface-based features were extracted following the Multi-center Epilepsy Lesion Detection pipeline. Features were normalized using intrasubject, interhemispheric, and intersubject z-scoring. A two-stage ML approach was applied: a vertexwise neural network classifier for lesional versus normal vertices using T1w/MRF/FLAIR features, followed by a clusterwise Random Undersampling Boosting classifier to suppress false positives (FPs) based on cluster size, prediction probabilities, and feature statistics. Leave-one-out cross-validation was performed at both stages.

RESULTS: Using T1w features, sensitivity was 70.4% with 11.6 FP clusters/patient and 4.1 in HCs. Adding MRF reduced FPs to 6.6 clusters/patient and 1.5 in HCs, with 68.2% sensitivity. Combining T1w, MRF, and FLAIR achieved 71.4% sensitivity, with 4.7 FPs/patient and 1.1 in HCs. Detection probabilities were significantly higher for true positive clusters than FPs (p < .001). Type II showed higher detection rates than non-type II. Magnetic resonance imaging (MRI)-positive patients showed higher detection rates and fewer FPs than MRI-negative patients. Seizure-free patients demonstrated higher detection rates than non-seizure-free patients. Subtyping accuracy was 80.8% for non-type II versus type II, and 68.4% for IIa versus IIb, although limited by small sample size. The transmantle sign was present in 61.5% of IIb and 40% of IIa cases.

SIGNIFICANCE: We developed an ML framework for FCD detection integrating SBM with clinical MRI and MRF. Advances include improved FP control and enhanced subtyping; selected model outputs may provide indicators of detection confidence and seizure outcome.

PMID:41066149 | DOI:10.1111/epi.18667