Search

Background: Artificial intelligence (AI)–enabled analysis of 12-lead ECGs may facilitate efficient estimation of incident atrial fibrillation (AF) risk. However, it remains unclear whether AI provides meaningful and generalizable improvement in predictive accuracy beyond clinical risk factors for AF. Methods: We trained a convolutional neural network (ECG-AI) to infer 5-year incident AF risk using 12-lead ECGs in patients receiving longitudinal primary care at Massachusetts General Hospital (MGH). We then fit 3 Cox proportional hazards models, composed of ECG-AI 5-year AF probability, CHARGE-AF clinical risk score (Cohorts for Heart and Aging in Genomic Epidemiology–Atrial Fibrillation), and terms for both ECG-AI and CHARGE-AF (CH-AI), respectively. We assessed model performance by calculating discrimination (area under the receiver operating characteristic curve) and calibration in an internal test set and 2 external test sets (Brigham and Women’s Hospital [BWH] and UK Biobank). Models were recalibrated to estimate 2-year AF risk in the UK Biobank given limited available follow-up. We used saliency mapping to identify ECG features most influential on ECG-AI risk predictions and assessed correlation between ECG-AI and CHARGE-AF linear predictors. Results: The training set comprised 45 770 individuals (age 55±17 years, 53% women, 2171 AF events) and the test sets comprised 83 162 individuals (age 59±13 years, 56% women, 2424 AF events). Area under the receiver operating characteristic curve was comparable using CHARGE-AF (MGH, 0.802 [95% CI, 0.767–0.836]; BWH, 0.752 [95% CI, 0.741–0.763]; UK Biobank, 0.732 [95% CI, 0.704–0.759]) and ECG-AI (MGH, 0.823 [95% CI, 0.790–0.856]; BWH, 0.747 [95% CI, 0.736–0.759]; UK Biobank, 0.705 [95% CI, 0.673–0.737]). Area under the receiver operating characteristic curve was highest using CH-AI (MGH, 0.838 [95% CI, 0.807 to 0.869]; BWH, 0.777 [95% CI, 0.766 to 0.788]; UK Biobank, 0.746 [95% CI, 0.716 to 0.776]). Calibration error was low using ECG-AI (MGH, 0.0212; BWH, 0.0129; UK Biobank, 0.0035) and CH-AI (MGH, 0.012; BWH, 0.0108; UK Biobank, 0.0001). In saliency analyses, the ECG P-wave had the greatest influence on AI model predictions. ECG-AI and CHARGE-AF linear predictors were correlated (Pearson r : MGH, 0.61; BWH, 0.66; UK Biobank, 0.41). Conclusions: AI-based analysis of 12-lead ECGs has similar predictive usefulness to a clinical risk factor model for incident AF and the approaches are complementary. ECG-AI may enable efficient quantification of future AF risk.

OBJECTIVE5-aminolevulinic acid (5-ALA)–induced protoporphyrin IX (PpIX) fluorescence is an effective surgical adjunct for the intraoperative identification of tumor tissue during resection of high-grade gliomas. The use of 5-ALA-induced PpIX fluorescence in glioblastoma (GBM) has been shown to double the extent of gross-total resection and 6-month progression-free survival. The heterogeneity of 5-ALA-induced PpIX fluorescence observed during surgery presents a technical and diagnostic challenge when utilizing this tool intraoperatively. While some regions show bright fluorescence after 5-ALA administration, other regions do not, despite that both regions of the tumor may be histopathologically indistinguishable. The authors examined the biological basis of this heterogeneity using computational methods.METHODSThe authors collected both fluorescent and nonfluorescent GBM specimens from a total of 14 patients undergoing surgery and examined their gene expression profiles.RESULTSIn this study, the authors found that the gene expression patterns characterizing fluorescent and nonfluorescent GBM surgical specimens were profoundly different and were associated with distinct cellular functions and different biological pathways. Nonfluorescent tumor tissue tended to resemble the neural subtype of GBM; meanwhile, fluorescent tumor tissue did not exhibit a prominent pattern corresponding to known subtypes of GBM. Consistent with this observation, neural GBM samples from The Cancer Genome Atlas database exhibited a significantly lower fluorescence score than nonneural GBM samples as determined by a fluorescence gene signature developed by the authors.CONCLUSIONSThese results provide a greater understanding regarding the biological basis of differential fluorescence observed intraoperatively and can provide a basis to identify novel strategies to maximize the effectiveness of fluorescence agents.

Purpose Surgical excision is currently recommended for all occurrences of atypical ductal hyperplasia (ADH) found on core needle biopsies for malignancy diagnoses and treatment of lesions. The excision of all ADH lesions may lead to overtreatment, which results in invasive surgeries for benign lesions in many women. A machine learning method to predict ADH upgrade may help clinicians and patients decide whether combined active surveillance and hormonal therapy is a reasonable alternative to surgical excision. Methods The following six machine learning models were developed to predict ADH upgrade from core needle biopsy: gradient-boosting trees, random forest, radial support vector machine (SVM), weighted K-nearest neighbors (KNN), logistic elastic net, and logistic regression. The study cohort consisted of 128 lesions from 124 women at a tertiary academic care center in New Hampshire who had ADH on core needle biopsy and who underwent an associated surgical excision from 2011 to 2017. Results The best-performing models were gradient-boosting trees (area under the curve [AUC], 68%; accuracy, 78%) and random forest (AUC, 67%; accuracy, 77%). The top five most important features that determined ADH upgrade were age at biopsy, lesion size, number of biopsies, needle gauge, and personal and family history of breast cancer. Using the random forest model, 98% of all malignancies would have been diagnosed through surgical biopsies, whereas 16% of unnecessary surgeries on benign lesions could have been avoided (ie, 87% sensitivity at 45% specificity). Conclusion These results add to the growing body of support for machine learning models as useful aids for clinicians and patients in decisions about the clinical management of ADH.

BackgroundElectronic health records (EHRs) promise to enable broad-ranging discovery with power exceeding that of conventional research cohort studies. However, research using EHR datasets may be subject to selection bias, which can be compounded by missing data, limiting the generalizability of derived insights.MethodsMass General Brigham (MGB) is a large New England-based healthcare network comprising seven tertiary care and community hospitals with associated outpatient practices. Within an MGB-based EHR warehouse of >3.5 million individuals with at least one ambulatory care visit, we approximated a community-based cohort study by selectively sampling individuals longitudinally attending primary care practices between 2001-2018 (n=520,868), which we named the Community Care Cohort Project (C3PO). We also utilized pre-trained deep natural language processing (NLP) models to recover vital signs (i.e., height, weight, and blood pressure) from unstructured notes in the EHR. We assessed the validity of C3PO by deploying established risk models including the Pooled Cohort Equations (PCE) and the Cohorts for Aging and Genomic Epidemiology Atrial Fibrillation (CHARGE-AF) score, and compared model performance in C3PO to that observed within typical EHR Convenience Samples which included all individuals from the same parent EHR with sufficient data to calculate each score but without a requirement for longitudinal primary care. All analyses were facilitated by the JEDI Extractive Data Infrastructure pipeline which we designed to efficiently aggregate EHR data within a unified framework conducive to regular updates.ResultsC3PO includes 520,868 individuals (mean age 48 years, 61% women, median follow-up 7.2 years, median primary care visits per individual 13). Estimated using reports, C3PO contains over 2.9 million electrocardiograms, 450,000 echocardiograms, 12,000 cardiac magnetic resonance images, and 75 million narrative notes. Using tabular data alone, 286,009 individuals (54.9%) had all vital signs available at baseline, which increased to 358,411 (68.8%) after NLP recovery (31% reduction in missingness). Among individuals with both NLP and tabular data available, NLP-extracted and tabular vital signs obtained on the same day were highly correlated (e.g., Pearson r range 0.95-0.99, pConclusionsIntentional sampling of individuals receiving regular ambulatory care and use of NLP to recover missing data have the potential to reduce bias in EHR research and maximize generalizability of insights.

The intersection of medicine and machine learning (ML) has the potential to transform healthcare. We describe how physiology, a foundational discipline of medical training and practice with a rich quantitative history, could serve as a starting point for the development of a common language between clinicians and ML experts, thereby accelerating real-world impact.

Colorectal cancer is the second leading cause of cancer-related deaths in the United States. Colorectal cancer risk can be effectively managed through early detection and removal of precancerous lesions, known as colorectal polyps, with routine colonoscopy screening. The current guidelines for colonoscopy screening and surveillance do not consider detailed clinical information and polyp characteristics from prior colonoscopies. Developing a colonoscopy surveillance plan based upon a patient’s personalized polyp recurrence risk is important for preventing the progression of colorectal cancer. To address this clinical need, in this paper, we proposed and developed a natural language processing and machine learning model to predict colorectal polyp recurrence risk using features derived from patient colonoscopy and pathology reports in electronic medical record systems. Colonoscopy records and the associated pathology reports from 952 patients in a tertiary academic care center in New Hampshire were obtained from 2011 to 2017. Polyp characteristics were extracted from these records using a natural language processing pipeline. The extracted features from these records along with other demographic and anthropometric information were used to develop and compare six machine learning models for their ability to predict polyp recurrence. Our evaluation of these models revealed a range of performance advantages, such as an area under the curve as high as 65%, and it further highlighted important features in predicting polyp recurrence from demographic and medical health record sources. Our predictive analysis highlights the potential of personalized risk modeling for colorectal cancer screening, which can reduce unnecessary screenings, healthcare costs, and psychological stress, while improving patient health outcomes.