Sina Sadeghi

Background: Type 2 diabetes (T2D) is a heterogeneous metabolic disorder composed of biologically and clinically distinct subtypes, i.e. severe insulin-deficient diabetes (SIDD), severe insulin-resistant diabetes (SIRD), mild obesity-related diabetes (MOD), and mild age-related diabetes (MARD), that associate with a differential risk for complications. Despite their promise for precision medicine, these clustering frameworks have been derived almost exclusively from outpatient cohorts of individuals with recent-onset T2D. However, no study has investigated the differential outcome of T2D subphenotypes in hospitalized patients.Methods: Using 8,114 adult patients admitted to a tertiary care unit (University of Leipzig Medical Center, Leipzig, Germany) for inpatient care from 2011–2022, we performed sex-stratified k-means clustering based on simple, easy-to-measure anthropometric and non-fasting clinical parameters, i.e. age, body mass index, glycated hemoglobin A1c, and triglyceride-to-high density lipoprotein cholesterol ratio (TG/HDL-C). Standardized inpatient outcomes included organ dysfunction (e.g. kidney replacement therapy and ventilation support), intensive care unit/intermediate care (ICU/IMC) admission, and in-hospital mortality. We applied machine learning-based approaches to predict cluster allocation and developed a publicly available online-calculator.Results: Clustering of T2D patients in hospitalized individuals based on routinely available easy-to-measure inpatient clinical markers successfully reproduces the well-established subphenotypes from population-based studies. Patients from the SIRD cluster show most pronounced, adverse phenotypes, including increased inflammation and impaired renal function at baseline. The SIRD cluster was associated with highest rates of mechanical ventilation and kidney replacement therapy during inpatient care, as well as ICU/IMC admission. Furthermore, SIRD and MARD patients were associated with significantly increased in-hospital mortality after adjustment for comorbidities. A multinomial logistic regression model predicted cluster membership with 92% accuracy. Conclusion: Our findings demonstrate that T2D subphenotypes can be identified using widely available, easily accessible clinical markers at hospital admission, enabling early risk stratification for prevalent comorbidities, as well as inpatient complications, including mortality. SIRD patients show the most adverse inpatient outcome.

The scarcity of high-quality clinical datasets and strict privacy regulations remain major barriers to developing robust predictive models in healthcare. Generative models offer a promising solution by producing synthetic data (SD) that replicate real data (RD) characteristics while mitigating privacy risks. This study introduces a structured framework for evaluating SD fidelity and utility using the Pima Indians Diabetes Dataset. Three tabular generative models, Conditional Tabular GAN (CTGAN), CopulaGAN, and Tabular Variational Autoencoder (TVAE), were assessed. Fidelity was quantified using the Pairwise Correlation Distance (PCD) and Wasserstein Distance (WSD), while utility was measured through binary classification performance. Optimized generative models generated high-fidelity SD with predictive accuracy comparable to RD. The results demonstrate that PCD and WSD are reliable indicators of SD quality and effective optimization targets for healthcare machine learning applications.

Objective: To assess the feasibility of a smartphone app delivering just‐in‐time adaptive interventions as an adjunct to cognitive‐behavioral therapy (CBT) adapted to binge‐eating disorder (BED), estimate its effects assuming superiority over CBT alone, and document safety and target engagement.Method: A single‐center, assessor‐blinded, parallel feasibility study randomized adults aged 18–65 years with full‐syndrome or subthreshold BED to smartphone‐supported CBT (SmartCBT) or standard CBT (DRKS00024597). Both arms received 16 individual 50‐min CBT sessions over 4 months. Assessments were conducted at baseline (T0), midtreatment (T1), posttreatment (T2), and 3‐month follow‐up (T3). Feasibility was determined regarding recruitment, attrition, dropout, adherence, assessment completion, app use, and acceptance. Further, eating disorder symptoms, mental and physical health, weight management behavior, safety, and target engagement (i.e., skill use) were assessed.Results: Over a 7‐month recruitment period, 28 of 50 eligible volunteers were included and randomized 1:1 to SmartCBT or CBT. In the modified intent‐to‐treat sample (N = 25; SmartCBT: 13, CBT: 12), the feasibility of SmartCBT was further supported regarding attrition, dropout, adherence, treatment completion, app use, and acceptance; however, assessment completion was moderate. Clinical improvements were found in both arms, but differential results were affected by baseline differences and moderate assessment completion in the SmartCBT arm. Safety was documented, and support for target engagement was found.Conclusions: This exploratory study provides evidence for the feasibility of app‐supported CBT for BED. With few procedural refinements, the protocol can be used in a confirmatory randomized‐controlled trial with long‐term follow‐up to evaluate efficacy and determine treatment mechanisms.Trial Registration: German Clinical Trials Register, https://www.drks.de, DRKS00024597

Defacing of brain magnetic resonance imaging (MRI) scans is a crucial process in medical imaging research aimed at preserving patient privacy while maintaining data integrity. However, existing defacing algorithms are prone to errors, potentially compromising patient anonymity. This paper investigates the feasibility and efficacy of automated quality assessment for defaced brain MRIs using machine learning (ML). Our findings demonstrate the promising capability of ML models in accurately distinguishing between properly and inadequately defaced MRI scans.