Dimitrios Zikos

Background/Objectives: Current clinical AI systems often report uncertainty as a static metric not tailored to the dynamic nature of clinical decision-making. To bridge the gap between computational variance and natural uncertainty in clinical practice, we propose the Contextual Uncertainty Reasoning for Artificial Intelligence (CURE-AI) framework. Methods: CURE-AI formalizes system uncertainty as a logical interaction between internal computational variance (the AI/ML pipeline) and external clinical ambiguity (the clinical workflow). The framework is constructed on five hierarchical levels of uncertainty resolution: (1) Clinical Uncertainty, (2) Computational Uncertainty, (3) Contextual Linkage, (4) Clinical Response, and (5) Human Outcome. Results: By mapping the framework to clinical workflows, we demonstrate that identical statistical uncertainty margins require different system recommendations depending on the patient context. CURE-AI translates computational variance into five actionable system behaviors: (1) inquiring for targeted missing data to resolve epistemic gaps, (2) accepting biological variability to prevent unnecessary testing, (3) scaling automated decisions or deferring to human judgment based on clinical risk, (4) suppressing AI output to adhere to mandated clinical protocols, and (5) recommending longitudinal patient reevaluation for prognostic uncertainty. Conclusions: CURE-AI recognizes AI uncertainty as an active workflow. By mapping computational variance to clinical action, the framework provides a conceptual blueprint for decision support systems that foster appropriate reliance and safeguard patient care.

Objective: Longitudinal trajectories are vital for clinical predictive modeling, yet many administrative healthcare datasets exhibit low temporal resolution, complicating chronological sequence extraction. We propose the TRACE (Temporal Resolution of Administrative Clinical Encounters) framework to extract, validate, and rank directed clinical pathways from nontemporal data, emphasizing the computational resolution of same-day ambiguity and the predictive value of “informative absences.” Methods: Using a dataset of 404,532 inpatient encounters from the 2024 CMS Limited Data Set mapped via CCSR, the TRACE framework first mitigates same-day documentation artifacts using Jaccard bundling (Jthresh ≥0.90). After establishing a 2% baseline support floor, candidate trajectories, including generated missing events (such as medical interventions), were evaluated using asymmetric probability and Markovian lift metrics and an empirical χ2 gate to prevent collider bias. Sequence nonrandomness was verified through a 1,000-iteration Monte Carlo permutation test with False Discovery Rate (FDR) control. Finally, to prove independent clinical value without relying on claim-level timestamps that do not represent actual clinical sequence, pathways were validated via a 5-stage ablation study utilizing L1-penalized regularization. Results: TRACE successfully extracted 463 trajectories. In the ablation study, the Standard Baseline (relying solely on the single raw features) achieved an initial PR-AUC of 0.2408. The Fully Engineered Hybrid model, com Longitudinal trajectories are vital for clinical predictive model stories, achieved the highest discrimination (ROC-AUC: 0.8093) and positive predictive value (PR-AUC: 0.2675). The L1 penalty compressed the 207-dimensional hybrid space down to 183 predictors, isolating 26 significant, multi-step clinical pathways predictive of in-hospital mortality. Conclusion: The TRACE framework extracts and empirically validates clinical trajectories from administrative data with low resolution or completely missing time information. By leveraging cross-sectional Boolean logic, informative absences, and predictive ablation, TRACE offers a scalable methodology for clinical sequence discovery.

This study evaluates diagnostic coding consistency across sequential hospitalizations using a longitudinal analysis of the CMS Inpatient Dataset with more than five million records. By transforming ICD-10-CM codes into CCS categories, the research tracks 9 chronic and 7 acute conditions using State Transition Matrices to identify patterns: Continuous, Fading, Intermittent Gap, and Late Onset. Significant diagnostic gaps were observed. Diabetes and Multiple Myeloma showed ‘Intermittent Gap’ rates of approximately 14%. Critically, for most chronic conditions, the ‘Intermittent Gap’ pattern correlated with the highest hospital Length of Stay, suggesting that missed documentation may obscure patient complexity. While some gaps stem from coding regulations or clinical transitions (such as remission or combination coding) the prevalence of intermittent patterns indicates systemic coding failures, with possible clinical coordination implications.

Objective: Longitudinal trajectories are vital for clinical predictive modeling, yet many administrative healthcare datasets exhibit low temporal resolution, complicating chronological sequence extraction. We propose the TRACE (Temporal Resolution of Administrative Clinical Encounters) framework to extract, validate, and rank directed clinical pathways from nontemporal data, emphasizing the computational resolution of same-day ambiguity and the predictive value of “informative absences.”Methods: Using a dataset of 404,532 inpatient encounters from the 2024 CMS Limited Data Set mapped via CCSR, the TRACE framework first mitigates same-day documentation artifacts using Jaccard bundling (Jthresh ≥0.90). After establishing a 2% baseline support floor, candidate trajectories, including generated missing events (such as medical interventions), were evaluated using asymmetric probability and Markovian lift metrics and an empirical χ2 gate to prevent collider bias. Sequence nonrandomness was verified through a 1,000-iteration Monte Carlo permutation test with False Discovery Rate (FDR) control. Finally, to prove independent clinical value without relying on claim-level timestamps that do not represent actual clinical sequence, pathways were validated via a 5-stage ablation study utilizing L1-penalized regularization.Results: TRACE successfully extracted 463 trajectories. In the ablation study, the Standard Baseline (relying solely on the single raw features) achieved an initial PR-AUC of 0.2408. The Fully Engineered Hybrid model, com Longitudinal trajectories are vital for clinical predictive model stories, achieved the highest discrimination (ROC-AUC: 0.8093) and positive predictive value (PR-AUC: 0.2675). The L1 penalty compressed the 207-dimensional hybrid space down to 183 predictors, isolating 26 significant, multi-step clinical pathways predictive of in-hospital mortality.Conclusion: The TRACE framework extracts and empirically validates clinical trajectories from administrative data with low resolution or completely missing time information. By leveraging cross-sectional Boolean logic, informative absences, and predictive ablation, TRACE offers a scalable methodology for clinical sequence discovery.

The fast growth of social media mining in health research has contributed to an invaluable but quite fragmented body of literature. As the amount of unstructured patient-reported data grows, traditional bibliometric analyses face methodological limitations, particularly regarding synonym fragmentation and arbitrary parameter selection. In their recent publication, “Thematic Mapping and Evolution of Social Media Mining in Health Research: Hybrid Bibliometric Synthesis,” Yang and Bohnet-Joschko attempt to address these flaws by introducing a semantic-structural (hybrid) bibliometric framework. This commentary evaluates the methodological innovations of their study and its departure from traditional syntactic keyword-matching tools. By combining citation-informed transformers (SPECTER2) and biomedical language models (PubMedBERT) and dimensionality reduction and density-based clustering, the authors created a reproducible pipeline. In their architecture, they start with foundational machine learning (statistical validity) before transitioning into large language models for qualitative synthesis. I will attempt to explain how this transition from syntactic mapping to semantic vector representation solves known challenges in evidence synthesis, naturally grouping conceptual synonyms without artificially forcing boundaries on the literature. Furthermore, I examine the practical implications of their temporal findings. Such real-time social media mining applications can be very useful for retrospective reporting and evaluating targeted public health interventions. While this pipeline offers high generalizability across disciplines, it also introduces a computational literacy barrier to some, and this re-emphasizes the need for data literacy for health professions. Ultimately, the study provides a transparent approach to informatics because mathematically validated frameworks are foundational for the future of evidence-driven public health policy and clinical decision-making.

Background/Objectives: Pressure ulcers are an important patient safety concern. While staging frameworks guide clinical management, the association between anatomical site, stage, and multiple PU presence and outcomes such as length of stay (LOS) and mortality in Medicare patients has not been fully characterized. The study objective is to examine the relationship between PU site, stage, and multiplicity and inpatient LOS and mortality among hospitalized Medicare patients. Methods: A cross-sectional study was conducted with 1,123,121 inpatient Medicare admissions from a 2019 Centers for Medicare and Medicaid Services (CMS) medical claims dataset. PUs were identified using ICD-10-CM codes classified by anatomical site and stage (1 through 4, unstageable, unspecified). Multiple regression models examined associations between PU characteristics and LOS and mortality, adjusting for age, sex, primary diagnosis, and hospital transfer. An analysis was conducted using SPSS version 29. Results: Stage 2 ulcers were the most common (28.6%), while unstageable or unspecified ulcers were frequent in the heels and head. The sacral region, buttocks, and heels were the most common anatomical sites. LOS gradually increased from Stage 1 (9.4 days) to Stage 4 (15.2 days). While the death rate did not increase consistently with stage, it was highest for upper back (14%), head (12.8%), and unspecified hip (12.8%) sites. Multiple regression was conducted to examine the association between PU locality, stage, and multiplicity and mortality (logistic regression) and LOS (linear regression). After controlling for patient demographics, admission, and clinical information, the regression results showed that the presence of multiple PUs, as well as anatomical sites of sacral, hip, head, buttock, and upper back ulcers, is associated with prolonged LOS and increased mortality. The presence of an advanced PU stage was found to be associated with prolonged LOS but not with inpatient mortality. Conclusions: The anatomical site and presence of multiple pressure ulcers were stronger predictors of prolonged hospital stay and increased mortality than ulcer stage alone. Certain ulcer sites, such as sacral, hip, head, buttock, and upper back, were independently associated with a higher risk of poor outcomes.

Background/Objectives: Pressure ulcers are an important patient safety healthcare concern. While staging frameworks guide clinical management, the association of the anatomical site, stage, and multiple PU presence, with outcomes such as length of stay (LOS) and mortality in Medicare patients has not been fully characterized. The study objective is to examine the relationship between PU site, stage, and multiplicity with inpatient LOS and mortality among hospitalized Medicare patients. Methods: This cross-sectional study was conducted with 1,123,121 inpatient Medicare admissions from a CMS medical claims dataset. PUs were identified using ICD-10-CM codes, classified by anatomical site and stage (1 though 4, unstageable, unspecified). Multivariate models examined associations of PU characteristics with LOS and mortality, adjusting for age, sex, primary diagnosis, and hospital transfer. Results: Overall, 3.7% of admissions included at least one PU (n=41,525). Stage 2 ulcers were the most common (28.6%), while unstageable or unspecified ulcers were frequent in heels and head. The sacral region was the most common anatomical site, followed by buttocks and heels. LOS gradually increased from Stage 1 (9.4 days) to Stage 4 (15.2 days). While death rate did not increase consistently with stage, it was highest for upper back (14%), head (12.8%), and unspecified hip (12.8%) sites. Multivariate analyses found sacral, hip, head, buttock, upper back ulcers and ulcers to be associated with prolonged LOS and increased mortality. Conclusions: PU anatomical site and multiple PU presence were stronger predictors of adverse outcomes than stage alone. These findings emphasize the importance of systematic documentation, early detection, and patient safety protocols to reduce the clinical and systemic burden of PUs.