Citations

Abstract Context Mandibular hypoplasia, deafness, progeroid features, and lipodystrophy (MDPL) syndrome is a rare, autosomal dominant disorder due to pathogenic heterozygous variants in POLD1. Clinical features of MDPL vary between patients; however, there is no previously reported genotype-phenotype association. Objective This work reports 14 new patients with lipodystrophy due to POLD1 variants and compares phenotypic differences between those with p.Ser605del and missense variants. Methods Genetic sequencing was performed on DNA of 14 patients for POLD1 variants, including exome (n = 10), genome (n = 1), and candidate gene (n = 3) sequencing. Comparisons of demographic, clinical features, and metabolic complications between carriers of POLD1 p.Ser605del and missense variants in our cases and those reported in the literature were made using the Fisher exact test for categorical variables and the t test for continuous variables. Results A total of 9 different POLD1 variants were identified in our patients, including 3 novel variants: p.Asp25Glufs*16, p.Arg507His, and p.Trp781Cys. Compared to individuals with missense variants (n = 15), those with the p.Ser605del (n = 26) POLD1 variant had significantly increased prevalence of mandibular hypoplasia (57% vs 100%, respectively; P = .015), small mouth (36% vs 100%, respectively; P = .015), crowded teeth (44% vs 91%, respectively; P = .046), and hypogonadism in male patients (0% vs 92%, respectively; P = .046). There were no differences in the prevalence of metabolic complications, such as diabetes, hypertriglyceridemia, and hepatic steatosis, in the two groups. Conclusion Individuals with the heterozygous POLD1 p.Ser605del variant had typical MDPL with more severe phenotype compared to those with missense variants with atypical MDPL.

Abstract Cytokine receptor–like factor 2 rearrangement (CRLF2r), an aberration causing a Philadelphia chromosome–like gene expression profile and a poor prognostic marker in B-cell acute lymphoblastic leukemia (B-ALL), endows relative chemotherapeutic resistance. Extramedullary disease (EMD), including both central nervous system (CNS) and non-CNS involvement, can also be associated with worse outcomes. The relationship between CRLF2r and EMD is unknown. Based on an initial observation of patients with relapsed/refractory (R/R) CRLF2r B-ALL and EMD, a single-center retrospective review was conducted to evaluate the incidence of EMD in patients with CRLF2r B-ALL. Across 177 children, adolescents, and young adults with R/R B-ALL, 21 (11.9%) harbored CRLF2r, of whom 19 (90.5%) had EMD at some point during their treatment course. In contrast, only 78 of the 156 patients (50%) without CRLF2r ever had EMD (P = .0003). In the CRLF2r cohort, 8 (38.1%) presented with EMD at diagnosis: 7 with CNS and 1 with non-CNS EMD. At relapse, 17 (80.9%) had EMD: 4 (19%) had CNS EMD, 9 (42.9%) had non-CNS EMD, and 4 (19%) had both. Isolated EMD was observed in 2 patients (11.8%) at first relapse and in 3 (17.6%) with multiple relapses. Given these findings, further study to evaluate this potential association between EMD and CRLF2r in R/R B-ALL is indicated.

Abstract Generative artificial intelligence (AI) is rapidly populating medical records with synthetic or partially AI-generated content, creating a feedback loop where future models are increasingly at risk of training on uncurated AI-generated data. However, the clinical consequences of this AI-generated data contamination remain unexplored. Here, we show that in the absence of mandatory human verification, this self-referential cycle drives a rapid erosion of pathological variability and diagnostic reliability of medical data at population scale. By analysing more than 800,000 synthetic data points across clinical text generation, vision–language reporting, and medical image synthesis, we find that models progressively converge toward generic phenotypes regardless of the model architectures. Specifically, rare but critical findings, including pneumothorax and effusions, vanish from the synthetic content generated by AI models, while demographic representations skew heavily toward middle-aged male phenotypes. Crucially, this degradation is masked by false diagnostic confidence. Models continue to issue reassuring reports while failing to detect life-threatening pathology, with false reassurance rates tripling to 40%. Blinded physician evaluation confirms that this decoupling of confidence and accuracy renders AI-generated documentation clinically useless after just two generations. We systematically evaluate three mitigation strategies that can be easily integrated into existing clinical workflows, finding that while synthetic volume scaling fails to prevent collapse, mixing real data with quality-aware filtering effectively preserves diversity. Ultimately, our results suggest that without policy-mandated human oversight, the deployment of generative AI threatens to degrade the very healthcare data ecosystems it relies upon.

Abstract Primary liver cancers, including hepatocellular carcinoma (HCC) and cholangiocarcinoma (CCA), arise from the neoplastic transformation of hepatocytes and cholangiocytes, respectively. Loss or downregulation of PTEN, a tumor suppressor negatively regulating the PI3K/AKT pathway, is frequently observed in CCA and HCC. Notably, PTEN mutations are observed at nearly twice the frequency in combined CCA-HCC tumors than either HCC or CCA alone. Using lineage-specific liver-targeted PTEN-deficient mouse models, we demonstrate that PTEN loss drives cellular dedifferentiation and tumorigenesis, a process that is critically dependent on AKT2. Mechanistically, PTEN deficiency induces activation of NOTCH and upregulation of transcriptional factor SOX9, which plays a central role in tumor cell transformation. In parallel, PTEN loss increases SMAD4 expression and sensitizes the tumor cells to TGFβ signaling, with TGFβ treatment repressing SOX9 expression in tumor cells lacking PTEN. Together, our study defined a critical role for PTEN-AKT2 signaling in maintaining liver epithelial lineage fidelity and revealed how its disruption promotes the conversion of mature hepatocytes or cholangiocytes into liver cancer stem-like cells (LCSCs). Furthermore, we identify a PTEN-dependent crosstalk between NOTCH and TGFβ pathways that governs liver tumor development. Together, this work provides mechanistic insight into lineage plasticity in liver cancer with implications for pathway-directed therapy.

Abstract Generative artificial intelligence (AI) is rapidly populating medical records with synthetic or partially AI-generated content, creating a feedback loop where future models are increasingly at risk of training on uncurated AI-generated data. However, the clinical consequences of this AI-generated data contamination remain unexplored. Here, we show that in the absence of mandatory human verification, this self-referential cycle drives a rapid erosion of pathological variability and diagnostic reliability of medical data at population scale. By analysing more than 800,000 synthetic data points across clinical text generation, vision–language reporting, and medical image synthesis, we find that models progressively converge toward generic phenotypes regardless of the model architectures. Specifically, rare but critical findings, including pneumothorax and effusions, vanish from the synthetic content generated by AI models, while demographic representations skew heavily toward middle-aged male phenotypes. Crucially, this degradation is masked by false diagnostic confidence. Models continue to issue reassuring reports while failing to detect life-threatening pathology, with false reassurance rates tripling to 40%. Blinded physician evaluation confirms that this decoupling of confidence and accuracy renders AI-generated documentation clinically useless after just two generations. We systematically evaluate three mitigation strategies that can be easily integrated into existing clinical workflows, finding that while synthetic volume scaling fails to prevent collapse, mixing real data with quality-aware filtering effectively preserves diversity. Ultimately, our results suggest that without policy-mandated human oversight, the deployment of generative AI threatens to degrade the very healthcare data ecosystems it relies upon.

Abstract Objective To measure the relative levels of signal and noise in expert diagnosis of epilepsy. Methods Twenty multinational epileptologists independently reviewed 50 vignettes of adult and pediatric patients presenting with suspected seizure(s) on two separate occasions with a ≥30‐day washout period. Experts provided a diagnosis of epilepsy or non‐epilepsy based on clinical information and, if requested, routine EEG and neuroimaging data. Cases had an established clinical diagnosis of epilepsy or non‐epilepsy based on capture of habitual paroxysmal events on video‐EEG or long‐term clinical follow‐up. Experts' judgments were analyzed to decompose variability into different sources: signal (objective differences between cases), level noise (experts' bias toward over/under‐diagnosis), pattern noise (experts' idiosyncratic reactions to specific case features), and occasion noise (inconsistency across occasions). Results The probability of an expert making a different diagnosis for a given case on two different occasions was 16%. The probability of two different experts making a different diagnosis for the same case was 26%. Signal (case “difficulty”) accounted for 66–69% of total variation, with 31–34% attributable to noise. Level noise was the largest contributor in the absence of EEG/neuroimaging results (23%), while pattern noise dominated when test results were available (24%). Occasion noise contributed relatively little (1%) but was still sufficient to cause diagnostic reversals in 16–22% between occasions. Significance The degree of noise in expert diagnosis of epilepsy is substantial, stemming primarily from physicians' idiosyncratic interpretations of case features and variable dispositions toward over‐ or under‐diagnosis. Strategies to improve reliability are needed, including standardized data collection protocols and structured decision algorithms. For “difficult cases,” where expert reliability and accuracy are lowest, our findings support current clinical practice which favors early referral for video‐EEG monitoring over reliance on diagnostic anchoring. This diagnostic pathway may become more accessible with advances in EEG technology (e.g., wearable devices) and artificial intelligence.