Iyer, Ravishankar K.

Abstract Cutaneous squamous cell carcinoma (cSCC) poses significant clinical challenges due to its rising incidence and potential for metastasis. Histopathologic risk stratification is further limited by substantial inter-observer variability. Unsupervised AI approaches based on content-based image retrieval offer scalable and interpretable decision support for diagnostic pathology. The objective of this study was to evaluate the use of image retrieval within histopathology atlases to stratify cSCC tumor differentiation from whole-slide images (WSIs), while comparing different patch selection and feature extraction strategies. This retrospective study included 552 archived WSIs comprising 385 well-differentiated, 102 moderately differentiated, and 66 poorly differentiated cases collected across Mayo Clinic sites in Arizona, Florida, and Minnesota. Image atlases were constructed using multiple patch aggregation strategies (Mosaic, Collage, and Montage) and deep learning encoders (KimiaNet, PathDino, and H-Optimus-0). A leave-one-WSI-out evaluation framework was used to assess differentiation classification performance using accuracy, specificity, sensitivity, and F1 score. Mosaic combined with KimiaNet achieved the highest Top-1 accuracy (74.9%) and specificity (92.6%), while Mosaic with H-Optimus-0 yielded the best Top-5 accuracy (79.0%) and macro-F1 score (62.6%). Collage combined with KimiaNet produced the highest Top-5 specificity (99.5%). The generalizability of the evaluated AI models varied across hospitals, reflecting differences in imaging protocols, staining practices, and patient populations. Overall, unsupervised image search and retrieval provides effective, annotation-free support for cSCC differentiation and has the potential to enhance dermatopathology workflows when appropriate combinations of patch selection and feature ex-traction methods are employed.

BACKGROUND There is a scarcity of artificial intelligence models trained on frozen pathology. One way to expand the clinical utility of models trained on permanent pathology is by applying them to frozen sections and fine-tune based on weaknesses. OBJECTIVE To qualitatively evaluate a deep learning model trained on permanent pathology to classify squamous cell carcinoma on Mohs surgery frozen sections to learn model shortcomings and inform retraining and fine-tuning. MATERIALS AND METHODS The authors trained a model for classification of tumor on 746 skin biopsy slides and tested it on 15 Mohs surgery frozen sections. The authors estimated performance metrics and compared the regions of interest generated by the model with the original H&E slides. RESULTS The model achieved an AUC-ROC of 0.985 and 0.796 for tumor classification in permanent pathology and in frozen sections, respectively. Regions of interest for frozen sections with scarce tumor areas were inaccurate, focusing on normal tissue for slides classified as false negative, or highlighting structures different from tumor (e.g., inflammation, muscle, and nerves) for slides classified as true positive. CONCLUSION Deep anatomical structures more commonly present in Mohs frozen pathology might represent data out-of-distribution for models trained on permanent pathology, potentially leading to inadequate model outputs.

Abstract Background: Primary sclerosing cholangitis (PSC) patients have a risk of developing cholangiocarcinoma (CCA). Establishing predictive models for CCA in PSC is important. Methods: In a large cohort of 1,459 PSC patients seen at Mayo Clinic (1993-2020), we quantified the impact of clinical/laboratory variables on CCA development using univariate and multivariate Cox models and predicted CCA using statistical and artificial intelligence approaches. We explored plasma bile acid (BA) levels’ predictive power of CCA (subset of 300 patients, BA cohort). Results: Eight significant risk factors (false discovery rate: 20%) were identified with univariate analysis; prolonged inflammatory bowel disease (IBD) was the most important one. IBD duration, PSC duration, and total bilirubin remained significant (p<0.05) with multivariate analysis. Clinical/laboratory variables predicted CCA with cross-validated C-indexes of 0.68–0.71 at different time points of disease, significantly better compared to commonly used PSC risk scores. Lower chenodeoxycholic acid, higher conjugated fraction of lithocholic acid and hyodeoxycholic acid, and higher ratio of cholic acid to chenodeoxycholic acid were predictive of CCA. BAs predicted CCA with a cross-validated C-index of 0.66 (std: 0.11, BA cohort), similar to clinical/laboratory variables (C-index=0.64, std: 0.11, BA cohort). Combining BAs with clinical/laboratory variables leads to the best average C-index of 0.67 (std: 0.13, BA cohort). Conclusions: In a large PSC cohort, we identified clinical and laboratory risk factors for CCA development and demonstrated the first artificial intelligence based predictive models that performed significantly better than common risk scores. More predictive data modalities are needed for clinical adoption of these models.

Abstract Motivation Telomeres are the repetitive sequences found at the ends of eukaryotic chromosomes and are often thought of as a ‘biological clock,’ with their average length shortening during division in most cells. In addition to their association with senescence, abnormal telomere lengths are well known to be associated with multiple cancers, short telomere syndromes and as risk factors for a broad range of diseases. While a majority of methods for measuring telomere length will report average lengths across all chromosomes, it is known that aberrations in specific chromosome arms are biomarkers for certain diseases. Due to their repetitive nature, characterizing telomeres at this resolution is prohibitive for short read sequencing approaches, and is challenging still even with longer reads. Results We present Telogator: a method for reporting chromosome-specific telomere length from long read sequencing data. We demonstrate Telogator’s sensitivity in detecting chromosome-specific telomere length in simulated data across a range of read lengths and error rates. Telogator is then applied to 10 germline samples, yielding a high correlation with short read methods in reporting average telomere length. In addition, we investigate common subtelomere rearrangements and identify the minimum read length required to anchor telomere/subtelomere boundaries in samples with these haplotypes. Availability and implementation Telogator is written in Python3 and is available at github.com/zstephens/telogator. Supplementary information Supplementary data are available at Bioinformatics online.

Abstract Objective Verbal memory dysfunction is common in focal, drug‐resistant epilepsy (DRE). Unfortunately, surgical removal of seizure‐generating brain tissue can be associated with further memory decline. Therefore, localization of both the circuits generating seizures and those underlying cognitive functions is critical in presurgical evaluations for patients who may be candidates for resective surgery. We used intracranial electroencephalographic (iEEG) recordings during a verbal memory task to investigate word encoding in focal epilepsy. We hypothesized that engagement in a memory task would exaggerate local iEEG feature differences between the seizure onset zone (SOZ) and neighboring tissue as compared to wakeful rest (“nontask”). Methods Ten participants undergoing presurgical iEEG evaluation for DRE performed a free recall verbal memory task. We evaluated three iEEG features in SOZ and non‐SOZ electrodes during successful word encoding and compared them with nontask recordings: interictal epileptiform spike (IES) rates, power in band (PIB), and relative entropy (REN; a functional connectivity measure). Results We found a complex pattern of PIB and REN changes in SOZ and non‐SOZ electrodes during successful word encoding compared to nontask. Successful word encoding was associated with a reduction in local electrographic functional connectivity (increased REN), which was most exaggerated in temporal lobe SOZ. The IES rates were reduced during task, but only in the non‐SOZ electrodes. Compared with nontask, REN features during task yielded marginal improvements in SOZ classification. Significance Previous studies have supported REN as a biomarker for epileptic brain. We show that REN differences between SOZ and non‐SOZ are enhanced during a verbal memory task. We also show that IESs are reduced during task in non‐SOZ, but not in SOZ. These findings support the hypothesis that SOZ and non‐SOZ respond differently to task and warrant further exploration into the use of cognitive tasks to identify functioning memory circuits and localize SOZ.

Abstract Routine scalp EEG is essential in the clinical diagnosis and management of epilepsy. However, a normal scalp EEG (based on expert visual review) recorded from a patient with epilepsy can cause delays in diagnosis and clinical care delivery. Here, we investigated whether normal EEGs might contain subtle electrophysiological clues of epilepsy. Specifically, we investigated (i) whether there are indicators of abnormal brain electrophysiology in normal EEGs of epilepsy patients, and (ii) whether such abnormalities are modulated by the side of the brain generating seizures in focal epilepsy. We analysed awake scalp EEG recordings of age-matched groups of 144 healthy individuals and 48 individuals with drug-resistant focal epilepsy who had normal scalp EEGs. After preprocessing, using a bipolar montage of eight channels, we extracted the fraction of spectral power in the alpha band (8–13 Hz) relative to a wide band of 0.5–40 Hz within 10-s windows. We analysed the extracted features for (i) the extent to which people with drug-resistant focal epilepsy differed from healthy subjects, and (ii) whether differences within the drug-resistant focal epilepsy patients were related to the hemisphere generating seizures. We then used those differences to classify whether an EEG is likely to have been recorded from a person with drug-resistant focal epilepsy, and if so, the epileptogenic hemisphere. Furthermore, we tested the significance of these differences while controlling for confounders, such as acquisition system, age and medications. We found that the fraction of alpha power is generally reduced (i) in drug-resistant focal epilepsy compared to healthy controls, and (ii) in right-handed drug-resistant focal epilepsy subjects with left hemispheric seizures compared to those with right hemispheric seizures, and that the differences are most prominent in the frontal and temporal regions. The fraction of alpha power yielded area under curve values of 0.83 in distinguishing drug-resistant focal epilepsy from healthy and 0.77 in identifying the epileptic hemisphere in drug-resistant focal epilepsy patients. Furthermore, our results suggest that the differences in alpha power are greater when compared with differences attributable to acquisition system differences, age and medications. Our findings support that EEG-based measures of normal brain function, such as the normalized spectral power of alpha activity, may help identify patients with epilepsy even when an EEG does not contain any epileptiform activity, recorded seizures or other abnormalities. Although alpha abnormalities are unlikely to be disease-specific, we propose that such abnormalities may provide a higher pre-test probability for epilepsy when an individual being screened for epilepsy has a normal EEG on visual assessment.

INTRODUCTION: Readmission and death in cirrhosis are common, expensive, and difficult to predict. Our aim was to evaluate the abilities of multiple artificial intelligence (AI) techniques to predict clinical outcomes based on variables collected at admission, during hospitalization, and at discharge. METHODS: We used the multicenter North American Consortium for the Study of End-Stage Liver Disease (NACSELD) cohort of cirrhotic inpatients who are followed up through 90-days postdischarge for readmission and death. We used statistical methods to select variables that are significant for readmission and death and trained 3 AI models, including logistic regression (LR), kernel support vector machine (SVM), and random forest classifiers (RFC), to predict readmission and death. We used the area under the receiver operating characteristic curve (AUC) from 10-fold crossvalidation for evaluation to compare sexes. Data were compared with model for end-stage liver disease (MELD) at discharge. RESULTS: We included 2,170 patients (57 ± 11 years, MELD 18 ± 7, 61% men, 79% White, and 8% Hispanic). The 30-day and 90-day readmission rates were 28% and 47%, respectively, and 13% died at 90 days. Prediction for 30-day readmission resulted in 0.60 AUC for all patients with RFC, 0.57 AUC with LR for women-only subpopulation, and 0.61 AUC with LR for men-only subpopulation. For 90-day readmission, the highest AUC was achieved with kernel SVM and RFC (AUC = 0.62). We observed higher predictive value when training models with only women (AUC = 0.68 LR) vs men (AUC = 0.62 kernel SVM). Prediction for death resulted in 0.67 AUC for all patients, 0.72 for women-only subpopulation, and 0.69 for men-only subpopulation, all with LR. MELD-Na model AUC was similar to those from the AI models. DISCUSSION: Despite using multiple AI techniques, it is difficult to predict 30- and 90-day readmissions and death in cirrhosis. AI model accuracies were equivalent to models generated using only MELD-Na scores. Additional biomarkers are needed to improve our predictive capability (See also the visual abstract at http://links.lww.com/AJG/B710).