Faculty Bibliography

The digital twin (DT) is a concept widely used in industry to create digital replicas of physical objects or systems. The dynamic, bi-directional link between the physical entity and its digital counterpart enables a real-time update of the digital entity. It can predict perturbations related to the physical object's function. The obvious applications of DTs in healthcare and medicine are extremely attractive prospects that have the potential to revolutionize patient diagnosis and treatment. However, challenges including technical obstacles, biological heterogeneity, and ethical considerations make it difficult to achieve the desired goal. Advances in multi-modal deep learning methods, embodied AI agents, and the metaverse may mitigate some difficulties. Here, we discuss the basic concepts underlying DTs, the requirements for implementing DTs in medicine, and their current and potential healthcare uses. We also provide our perspective on five hallmarks for a healthcare DT system to advance research in this field.

INTRODUCTION/BACKGROUND:Hematoma expansion (HE) in patients with intracerebral hemorrhage (ICH) is a key predictor of poor prognosis and potentially amenable to treatment. This study aimed to build a classification model to predict HE in patients with ICH using deep learning algorithms without using advanced radiological features. METHODS:Data from the ATACH-2 trial (Antihypertensive Treatment of Acute Cerebral Hemorrhage) was utilized. Variables included in the models were chosen as per literature consensus on salient variables associated with HE. HE was defined as increase in either >33% or 6 mL in hematoma volume in the first 24 h. Multiple machine learning algorithms were employed using iterative feature selection and outcome balancing methods. 70% of patients were used for training and 30% for internal validation. We compared the ML models to a logistic regression model and calculated AUC, accuracy, sensitivity and specificity for the internal validation models respective models. RESULTS:[5.03-18.17] and 25.2% had HE. The best performing model across all feature selection groups and sampling cohorts was using an artificial neural network (ANN) for HE in the testing cohort with AUC 0.702 [95% CI, 0.631-0.774] with 8 hidden layer nodes The traditional logistic regression yielded AUC 0.658 [95% CI, 0.641-0.675]. All other models performed with less accuracy and lower AUC. Initial hematoma volume, time to initial CT head, and initial SBP emerged as most relevant variables across all best performing models. CONCLUSION/CONCLUSIONS:We developed multiple ML algorithms to predict HE with the ANN classifying the best without advanced radiographic features, although the AUC was only modestly better than other models. A larger, more heterogenous dataset is needed to further build and better generalize the models.

The most widely used fluorophore in glioma-resection surgery, 5-aminolevulinic acid (5-ALA), is thought to cause the selective accumulation of fluorescent protoporphyrin IX (PpIX) in tumour cells. Here we show that the clinical detection of PpIX can be improved via a microscope that performs paired stimulated Raman histology and two-photon excitation fluorescence microscopy (TPEF). We validated the technique in fresh tumour specimens from 115 patients with high-grade gliomas across four medical institutions. We found a weak negative correlation between tissue cellularity and the fluorescence intensity of PpIX across all imaged specimens. Semi-supervised clustering of the TPEF images revealed five distinct patterns of PpIX fluorescence, and spatial transcriptomic analyses of the imaged tissue showed that myeloid cells predominate in areas where PpIX accumulates in the intracellular space. Further analysis of external spatially resolved metabolomics, transcriptomics and RNA-sequencing datasets from glioblastoma specimens confirmed that myeloid cells preferentially accumulate and metabolize PpIX. Our findings question 5-ALA-induced fluorescence in glioma cells and show how 5-ALA and TPEF imaging can provide a window into the immune microenvironment of gliomas.

STUDY DESIGN/METHODS:A retrospective cohort study. OBJECTIVE:The purpose of this study is to develop a machine learning algorithm to predict nonhome discharge after cervical spine surgery that is validated and usable on a national scale to ensure generalizability and elucidate candidate drivers for prediction. SUMMARY OF BACKGROUND DATA/BACKGROUND:Excessive length of hospital stay can be attributed to delays in postoperative referrals to intermediate care rehabilitation centers or skilled nursing facilities. Accurate preoperative prediction of patients who may require access to these resources can facilitate a more efficient referral and discharge process, thereby reducing hospital and patient costs in addition to minimizing the risk of hospital-acquired complications. METHODS:Electronic medical records were retrospectively reviewed from a single-center data warehouse (SCDW) to identify patients undergoing cervical spine surgeries between 2008 and 2019 for machine learning algorithm development and internal validation. The National Inpatient Sample (NIS) database was queried to identify cervical spine fusion surgeries between 2009 and 2017 for external validation of algorithm performance. Gradient-boosted trees were constructed to predict nonhome discharge across patient cohorts. The area under the receiver operating characteristic curve (AUROC) was used to measure model performance. SHAP values were used to identify nonlinear risk factors for nonhome discharge and to interpret algorithm predictions. RESULTS:A total of 3523 cases of cervical spine fusion surgeries were included from the SCDW data set, and 311,582 cases were isolated from NIS. The model demonstrated robust prediction of nonhome discharge across all cohorts, achieving an area under the receiver operating characteristic curve of 0.87 (SD=0.01) on both the SCDW and nationwide NIS test sets. Anterior approach only, age, elective admission status, Medicare insurance status, and total Elixhauser Comorbidity Index score were the most important predictors of discharge destination. CONCLUSIONS:Machine learning algorithms reliably predict nonhome discharge across single-center and national cohorts and identify preoperative features of importance following cervical spine fusion surgery.

OBJECTIVE:To examine the usefulness of carotid web (CW), carotid bifurcation and their combined angioarchitectural measurements in assessing stroke risk. METHODS:Anatomic data on the internal carotid artery (ICA), common carotid artery (CCA), and the CW were gathered as part of a retrospective study from symptomatic (stroke) and asymptomatic (nonstroke) patients with CW. We built a model of stroke risk using principal-component analysis, Firth regression trained with 5-fold cross-validation, and heuristic binary cutoffs based on the Minimal Description Length principle. RESULTS:The study included 22 patients, with a mean age of 55.9 ± 12.8 years; 72.9% were female. Eleven patients experienced an ischemic stroke. The first 2 principal components distinguished between patients with stroke and patients without stroke. The model showed that ICA-pouch tip angle (P = 0.036), CCA-pouch tip angle (P = 0.036), ICA web-pouch angle (P = 0.036), and CCA web-pouch angle (P = 0.036) are the most important features associated with stroke risk. Conversely, CCA and ICA anatomy (diameter and angle) were not found to be risk factors. CONCLUSIONS:This pilot study shows that using data from computed tomography angiography, carotid bifurcation, and CW angioarchitecture may be used to assess stroke risk, allowing physicians to tailor care for each patient according to risk stratification.

TOPIC/UNASSIGNED:This scoping review summarizes artificial intelligence (AI) reporting in ophthalmology literature in respect to model development and validation. We characterize the state of transparency in reporting of studies prospectively validating models for disease classification. CLINICAL RELEVANCE/UNASSIGNED:Understanding what elements authors currently describe regarding their AI models may aid in the future standardization of reporting. This review highlights the need for transparency to facilitate the critical appraisal of models prior to clinical implementation, to minimize bias and inappropriate use. Transparent reporting can improve effective and equitable use in clinical settings. METHODS/UNASSIGNED:Eligible articles (as of January 2022) from PubMed, Embase, Web of Science, and CINAHL were independently screened by 2 reviewers. All observational and clinical trial studies evaluating the performance of an AI model for disease classification of ophthalmic conditions were included. Studies were evaluated for reporting of parameters derived from reporting guidelines (CONSORT-AI, MI-CLAIM) and our previously published editorial on model cards. The reporting of these factors, which included basic model and dataset details (source, demographics), and prospective validation outcomes, were summarized. RESULTS/UNASSIGNED:Thirty-seven prospective validation studies were included in the scoping review. Eleven additional associated training and/or retrospective validation studies were included if this information could not be determined from the primary articles. These 37 studies validated 27 unique AI models; multiple studies evaluated the same algorithms (EyeArt, IDx-DR, and Medios AI). Details of model development were variably reported; 18 of 27 models described training dataset annotation and 10 of 27 studies reported training data distribution. Demographic information of training data was rarely reported; 7 of the 27 unique models reported age and gender and only 2 reported race and/or ethnicity. At the level of prospective clinical validation, age and gender of populations was more consistently reported (29 and 28 of 37 studies, respectively), but only 9 studies reported race and/or ethnicity data. Scope of use was difficult to discern for the majority of models. Fifteen studies did not state or imply primary users. CONCLUSION/UNASSIGNED:Our scoping review demonstrates variable reporting of information related to both model development and validation. The intention of our study was not to assess the quality of the factors we examined, but to characterize what information is, and is not, regularly reported. Our results suggest the need for greater transparency in the reporting of information necessary to determine the appropriateness and fairness of these tools prior to clinical use. FINANCIAL DISCLOSURES/UNASSIGNED:Proprietary or commercial disclosure may be found in the Footnotes and Disclosures at the end of this article.

BACKGROUND AND OBJECTIVES/OBJECTIVE:Clinical registries are critical for modern surgery and underpin outcomes research, device monitoring, and trial development. However, existing approaches to registry construction are labor-intensive, costly, and prone to manual error. Natural language processing techniques combined with electronic health record (EHR) data sets can theoretically automate the construction and maintenance of registries. Our aim was to automate the generation of a spine surgery registry at an academic medical center using regular expression (regex) classifiers developed by neurosurgeons to combine domain expertise with interpretable algorithms. METHODS:We used a Hadoop data lake consisting of all the information generated by an academic medical center. Using this database and structured query language queries, we retrieved every operative note written in the department of neurosurgery since our transition to EHR. Notes were parsed using regex classifiers and compared with a random subset of 100 manually reviewed notes. RESULTS:A total of 31 502 operative cases were downloaded and processed using regex classifiers. The codebase required 5 days of development, 3 weeks of validation, and less than 1 hour for the software to generate the autoregistry. Regex classifiers had an average accuracy of 98.86% at identifying both spinal procedures and the relevant vertebral levels, and it correctly identified the entire list of defined surgical procedures in 89% of patients. We were able to identify patients who required additional operations within 30 days to monitor outcomes and quality metrics. CONCLUSION/CONCLUSIONS:This study demonstrates the feasibility of automatically generating a spine registry using the EHR and an interpretable, customizable natural language processing algorithm which may reduce pitfalls associated with manual registry development and facilitate rapid clinical research.

BACKGROUND AND OBJECTIVES/OBJECTIVE:Advances in targeted therapies and wider application of stereotactic radiosurgery (SRS) have redefined outcomes of patients with brain metastases. Under modern treatment paradigms, there remains limited characterization of which aspects of disease drive demise and in what frequencies. This study aims to characterize the primary causes of terminal decline and evaluate differences in underlying intracranial tumor dynamics in patients with metastatic brain cancer. These fundamental details may help guide management, patient counseling, and research priorities. METHODS:Using NYUMets-Brain-the largest, longitudinal, real-world, open data set of patients with brain metastases-patients treated at New York University Langone Health between 2012 and 2021 with SRS were evaluated. A review of electronic health records allowed for the determination of a primary cause of death in patients who died during the study period. Causes were classified in mutually exclusive, but collectively exhaustive, categories. Multilevel models evaluated for differences in dynamics of intracranial tumors, including changes in volume and number. RESULTS:Of 439 patients with end-of-life data, 73.1% died secondary to systemic disease, 10.3% died secondary to central nervous system (CNS) disease, and 16.6% died because of other causes. CNS deaths were driven by acute increases in intracranial pressure (11%), development of focal neurological deficits (18%), treatment-resistant seizures (11%), and global decline driven by increased intracranial tumor burden (60%). Rate of influx of new intracranial tumors was almost twice as high in patients who died compared with those who survived (P < .001), but there was no difference in rates of volume change per intracranial tumor (P = .95). CONCLUSION/CONCLUSIONS:Most patients with brain metastases die secondary to systemic disease progression. For patients who die because of neurological disease, tumor dynamics and cause of death mechanisms indicate that the primary driver of decline for many may be unchecked systemic disease with unrelenting spread of new tumors to the CNS rather than failure of local growth control.

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 deployment^1-3. 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 processing^4,5 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.

Artificial intelligence and machine learning applications are becoming increasingly popular in health care and medical devices. The development of accurate machine learning algorithms requires large quantities of good and diverse data. This poses a challenge in health care because of the sensitive nature of sharing patient data. Decentralized algorithms through federated learning avoid data aggregation. In this paper we give an overview of federated learning, current examples in healthcare and ophthalmology, challenges, and next steps.