The exponential growth of epidemiological data from distributed healthcare systems presents unprecedented opportunities and challenges for communicable disease surveillance. This paper presents a comprehensive cloud-based analytics framework leveraging Apache Spark for distributed processing of 1.8 million epidemiological records to enable real-time communicable disease prediction using sparse clinical inputs. We propose a privacy-preserving architecture that integrates local language models via Ollama at the edge layer for initial data structuring, ensuring patient privacy and data sovereignty before cloud transmission. The core predictive component employs a Random Forest Classifier trained on 86 clinical features spanning demographic, physiological, hematological, immunological, and symptomatic domains. Our framework achieves 89.0
In this research, structural interaction analysis of Perylene Diimide (PDI) core dendrimers is conducted over four therapeutic and imaging classes. Core topology was mathematically encoded using topological descriptors, which include Leap Zagreb indices and polymer path Wiener polarity index. Drug generation encapsulation suitability was quantified through global feature normalization and standardized Euclidean distance, yielding compatibility scores used to rank dendrimer generations ( D_0 - D_4 ) for each drug without altering drug order or identity. To assess the stability and systematicity of generation suitability trends, intergenerational behavior was evaluated using monotonic rank association via Spearman’s rank correlation and global compatibility magnitude through relative changes in mean compatibility. To determine which topological indices most effectively explain specific physicochemical properties of drugs, that are, molecular weight, surface polarity, hydrogen bonding capacity, rotatable bond flexibility, π -complexity, and structural intricacy, Gaussian Process Regression (GPR) models were developed within a Quantitative Structure-Property Relationship (QSPR) framework using a descriptor space comprising four therapeutic drug-related chemical indices. Model outputs were carefully analysed to evaluate descriptor property relevance using prediction error metrics and correlation statistics, allowing us to determine the most informative and relevant topological index for each chemical attribute. These results demonstrate that the employed topological descriptors provide a systematic graph-based representation for generation wise compatibility and exhibit property-dependent relationships with diverse physicochemical characteristics.
Pediatric Attention Deficit Hyperactivity Disorder (ADHD) presents substantial challenges for digital health interventions due to its symptomatic heterogeneity and developmental specificity. Recent advances in Large Language Models have enabled significant improvements in natural language understanding and generation, yet most existing therapeutic chatbots rely on monolithic architectures that do not address the distinct requirements of attention deficits, hyperactivity, executive dysfunction, emotional dysregulation, and working memory limitations. This paper introduces ADHD-Agents, an architecture integrating five specialized coaching agents within a retrieval-augmented generation framework grounded in carefully curated knowledge sources. The knowledge base comprises 500 child-oriented question–answer cards organized into ten thematic categories, informed by healthcare professionals and reviewed by the project team, on which the agents ground their advice, and an indexed corpus of ADHD research papers, prepared as an extension to enrich the agents’ sources. A workflow coordinates agent selection and conversation transfers by combining retrieval-based voting with LLM-based intent classification. Emotional cues can trigger the suggestion of a complementary specialist, while a dedicated safety mechanism handles sensitive disclosures. A dual-memory system keeps track of the ongoing conversation and builds, across sessions, a profile of each child, so that suggestions and encouragement are adapted to the difficulties expressed and to the child’s feedback on previous strategies. The architecture also incorporates privacy-oriented measures and parent-mediated synthesis reports, and is designed to integrate therapeutic gaming modules in the future. A preliminary experimental validation using a synthetic dataset of test scenarios examined the system’s core technical functionalities, including information retrieval, agent routing, safety handling, session memory, and cross-session personalization. This evaluation provides an initial assessment of the architecture’s functional feasibility, while its therapeutic effectiveness and suitability for real-world use remain to be established through further studies.
Synthetic instruction data offer a scalable strategy for adapting large language models to computational pathology, where access to privacy-compliant, expert-annotated clinical text remains limited. However, the effect of synthetic-data generator selection on downstream model performance has not been systematically evaluated for pathology-report question answering. We compared six Google Gemini generator configurations, including Gemini 2 Flash, Gemini 2 Flash Lite, Gemini 2.5 Flash, Gemini 2.5 Flash Latest, Gemini 2.5 Flash Lite, and Gemini 3 Flash, for generating report-grounded question–answer pairs from 1,015 TCGA breast invasive carcinoma pathology reports. Each generator was prompted to generate 10 question–answer pairs per report, and the resulting datasets were subsequently used to fine-tune Meta-Llama-3.1-8B using 4-bit quantization and Quantized Low-Rank Adaptation. To reduce dependence on a single train–validation split, the complete fine-tuning and evaluation workflow was repeated using three random split seeds (42, 56, and 1093), while keeping the fine-tuning seed and training settings fixed. Performance was assessed using validation loss, perplexity, ROUGE, and BERTScore. In addition, a blinded, source-grounded factual-consistency assessment was conducted on 300 stratified question–answer pairs, with all automatically flagged cases manually reviewed against the original pathology reports. Gemini 2 Flash consistently achieved the lowest validation loss (60.71 ± 0.58) and perplexity (1.8350 ± 0.0106), indicating the strongest likelihood-based fit. Gemini 2.5 Flash achieved the highest ROUGE-1 (56.75 ± 0.82), ROUGE-2 (38.52 ± 1.03), ROUGE-L (53.11 ± 1.05), and BERTScore F1 (93.80 ± 0.07). These rankings remained unchanged across all three random train/validation partitions, and across-split variability was small. Among 297 evaluable answers in the source-grounded audit, 99.7
Multi-center studies are often necessary in rare pediatric conditions, but direct pooling of patient-level data is constrained by privacy, governance, and hospital-integration barriers. We examine whether routine clinical statistics can be reproduced over encrypted data for a fixed hospital consortium. We implemented Pearson correlation, Wilcoxon rank-sum, and chi-squared tests with threshold Cheon–Kim–Kim–Song (CKKS) homomorphic encryption in the open-source OpenFHE library. Evaluation used a pediatric craniospinal irradiation use case focused on hematologic toxicity; the multi-center scenario was emulated by partitioning the original 51-patient cohort into two and three site-specific subsets, plus synthetic cohorts for scaling. We also define an operational governance model covering role separation, key-share custody, query approval, auditable decryption, and hospital-side extract–transform–load (ETL) and encryption boundaries. Encrypted Pearson correlation and chi-squared stayed within 0.001
