Physicians make critical time-constrained decisions every day. Clinical predictive models can help physicians and administrators make decisions by forecasting clinical and operational events. Existing structured data-based clinical predictive models have limited use in everyday practice owing to complexity in data processing, as well as model development and deployment1–3. Here we show that unstructured clinical notes from the electronic health record can enable the training of clinical language models, which can be used as all-purpose clinical predictive engines with low-resistance development and deployment. Our approach leverages recent advances in natural language processing4,5 to train a large language model for medical language (NYUTron) and subsequently fine-tune it across a wide range of clinical and operational predictive tasks. We evaluated our approach within our health system for five such tasks: 30-day all-cause readmission prediction, in-hospital mortality prediction, comorbidity index prediction, length of stay prediction, and insurance denial prediction. We show that NYUTron has an area under the curve (AUC) of 78.7–94.9%, with an improvement of 5.36–14.7% in the AUC compared with traditional models. We additionally demonstrate the benefits of pretraining with clinical text, the potential for increasing generalizability to different sites through fine-tuning and the full deployment of our system in a prospective, single-arm trial. These results show the potential for using clinical language models in medicine to read alongside physicians and provide guidance at the point of care.

Background: Tumor programmed death-ligand 1 (PD-L1) status is useful in determining which patients may benefit from programmed death-1 (PD-1)/PD-L1 inhibitors. However, little is known about the association between PD-L1 status and tumor histopathological patterns. Using deep learning, we predicted PD-L1 status from hematoxylin and eosin (H and E) whole-slide images (WSIs) of nonsmall cell lung cancer (NSCLC) tumor samples. Materials and Methods: One hundred and thirty NSCLC patients were randomly assigned to training ( n = 48) or test ( n = 82) cohorts. A pair of H and E and PD-L1-immunostained WSIs was obtained for each patient. A pathologist annotated PD-L1 positive and negative tumor regions on the training samples using immunostained WSIs for reference. From the H and E WSIs, over 145,000 training tiles were generated and used to train a multi-field-of-view deep learning model with a residual neural network backbone. Results: The trained model accurately predicted tumor PD-L1 status on the held-out test cohort of H and E WSIs, which was balanced for PD-L1 status (area under the receiver operating characteristic curve [AUC] =0.80, P << 0.01). The model remained effective over a range of PD-L1 cutoff thresholds (AUC = 0.67–0.81, P ≤ 0.01) and when different proportions of the labels were randomly shuffled to simulate interpathologist disagreement (AUC = 0.63–0.77, P ≤ 0.03). Conclusions: A robust deep learning model was developed to predict tumor PD-L1 status from H and E WSIs in NSCLC. These results suggest that PD-L1 expression is correlated with the morphological features of the tumor microenvironment.

OBJECTIVEUnplanned preventable hospital readmissions within 30 days are a great burden to patients and the healthcare system. With an estimated $41.3 billion spent yearly, reducing such readmission rates is of the utmost importance. With the widespread adoption of big data and machine learning, clinicians can use these analytical tools to understand these complex relationships and find predictive factors that can be generalized to future patients. The object of this study was to assess the efficacy of a machine learning algorithm in the prediction of 30-day hospital readmission after posterior spinal fusion surgery.METHODSThe authors analyzed the distribution of National Surgical Quality Improvement Program (NSQIP) posterior lumbar fusions from 2011 to 2016 by using machine learning techniques to create a model predictive of hospital readmissions. A deep neural network was trained using 177 unique input variables. The model was trained and tested using cross-validation, in which the data were randomly partitioned into training (n = 17,448 [75%]) and testing (n = 5816 [25%]) data sets. In training, the 17,448 training cases were fed through a series of 7 layers, each with varying degrees of forward and backward communicating nodes (neurons).RESULTSMean and median positive predictive values were 78.5% and 78.0%, respectively. Mean and median negative predictive values were both 97%, respectively. Mean and median areas under the curve for the model were 0.812 and 0.810, respectively. The five most heavily weighted inputs were (in order of importance) return to the operating room, septic shock, superficial surgical site infection, sepsis, and being on a ventilator for > 48 hours.CONCLUSIONSMachine learning and artificial intelligence are powerful tools with the ability to improve understanding of predictive metrics in clinical spine surgery. The authors’ model was able to predict those patients who would not require readmission. Similarly, the majority of predicted readmissions (up to 60%) were predicted by the model while retaining a 0% false-positive rate. Such findings suggest a possible need for reevaluation of the current Hospital Readmissions Reduction Program penalties in spine surgery.