Faculty Bibliography

The fundamental challenge in machine learning is ensuring that trained models generalize well to unseen data. We developed a general technique for ameliorating the effect of dataset shift using generative adversarial networks (GANs) on a dataset of 149,298 handwritten digits and dataset of 868,549 chest radiographs obtained from four academic medical centers. Efficacy was assessed by comparing area under the curve (AUC) pre- and post-adaptation. On the digit recognition task, the baseline CNN achieved an average internal test AUC of 99.87% (95% CI, 99.87-99.87%), which decreased to an average external test AUC of 91.85% (95% CI, 91.82-91.88%), with an average salvage of 35% from baseline upon adaptation. On the lung pathology classification task, the baseline CNN achieved an average internal test AUC of 78.07% (95% CI, 77.97-78.17%) and an average external test AUC of 71.43% (95% CI, 71.32-71.60%), with a salvage of 25% from baseline upon adaptation. Adversarial domain adaptation leads to improved model performance on radiographic data derived from multiple out-of-sample healthcare populations. This work can be applied to other medical imaging domains to help shape the deployment toolkit of machine learning in medicine.

Federated learning (FL) is a method used for training artificial intelligence models with data from multiple sources while maintaining data anonymity, thus removing many barriers to data sharing. Here we used data from 20 institutes across the globe to train a FL model, called EXAM (electronic medical record (EMR) chest X-ray AI model), that predicts the future oxygen requirements of symptomatic patients with COVID-19 using inputs of vital signs, laboratory data and chest X-rays. EXAM achieved an average area under the curve (AUC) >0.92 for predicting outcomes at 24 and 72 h from the time of initial presentation to the emergency room, and it provided 16% improvement in average AUC measured across all participating sites and an average increase in generalizability of 38% when compared with models trained at a single site using that site's data. For prediction of mechanical ventilation treatment or death at 24 h at the largest independent test site, EXAM achieved a sensitivity of 0.950 and specificity of 0.882. In this study, FL facilitated rapid data science collaboration without data exchange and generated a model that generalized across heterogeneous, unharmonized datasets for prediction of clinical outcomes in patients with COVID-19, setting the stage for the broader use of FL in healthcare.

Real-time seizure detection is a resource intensive process as it requires continuous monitoring of patients on stereoelectroencephalography. This study improves real-time seizure detection in drug resistant epilepsy (DRE) patients by developing patient-specific deep learning models that utilize a novel self-supervised dynamic thresholding approach. Deep neural networks were constructed on over 2000 h of high-resolution, multichannel SEEG and video recordings from 14 DRE patients. Consensus labels from a panel of epileptologists were used to evaluate model efficacy. Self-supervised dynamic thresholding exhibited improvements in positive predictive value (PPV; difference: 39.0%; 95% CI 4.5-73.5%; Wilcoxon-Mann-Whitney test; N = 14; p = 0.03) with similar sensitivity (difference: 14.3%; 95% CI - 21.7 to 50.3%; Wilcoxon-Mann-Whitney test; N = 14; p = 0.42) compared to static thresholds. In some models, training on as little as 10 min of SEEG data yielded robust detection. Cross-testing experiments reduced PPV (difference: 56.5%; 95% CI 25.8-87.3%; Wilcoxon-Mann-Whitney test; N = 14; p = 0.002), while multimodal detection significantly improved sensitivity (difference: 25.0%; 95% CI 0.2-49.9%; Wilcoxon-Mann-Whitney test; N = 14; p < 0.05). Self-supervised dynamic thresholding improved the efficacy of real-time seizure predictions. Multimodal models demonstrated potential to improve detection. These findings are promising for future deployment in epilepsy monitoring units to enable real-time seizure detection without annotated data and only minimal training time in individual patients.

Purpose/UNASSIGNED:To train a deep learning classification algorithm to predict chest radiograph severity scores and clinical outcomes in patients with coronavirus disease 2019 (COVID-19). Materials and Methods/UNASSIGNED:= 110) populations. Bootstrapping was used to compute CIs. Results/UNASSIGNED:The model trained on the chest radiograph severity score produced the following areas under the receiver operating characteristic curves (AUCs): 0.80 (95% CI: 0.73, 0.88) for the chest radiograph severity score, 0.76 (95% CI: 0.68, 0.84) for admission, 0.66 (95% CI: 0.56, 0.75) for intubation, and 0.59 (95% CI: 0.49, 0.69) for death. The model trained on clinical variables produced an AUC of 0.64 (95% CI: 0.55, 0.73) for intubation and an AUC of 0.59 (95% CI: 0.50, 0.68) for death. Combining chest radiography and clinical variables increased the AUC of intubation and death to 0.88 (95% CI: 0.79, 0.96) and 0.82 (95% CI: 0.72, 0.91), respectively. Conclusion/UNASSIGNED:© RSNA, 2020.

BACKGROUND:Understanding the interactions between variables that predict prolonged hospital length of stay (LOS) following spine surgery can help uncover drivers of this risk in patients. This study utilized a novel game-theory-based approach to develop explainable machine learning models to understand such interactions in a large cohort of patients treated with spine surgery. METHODS:Of 11,150 patients who underwent surgery for degenerative spine conditions at a single institution, 3,310 (29.7%) were characterized as having prolonged LOS. Machine learning models predicting LOS were built for each patient. Shapley additive explanation (SHAP) values were calculated for each patient model to quantify the importance of features and variable interaction effects. RESULTS:Models using features identified by SHAP values were highly predictive of prolonged LOS risk (mean C-statistic = 0.87). Feature importance analysis revealed that prolonged LOS risk is multifactorial. Non-elective admission produced elevated SHAP values, indicating a clear, strong risk of prolonged LOS. In contrast, intraoperative and sociodemographic factors displayed bidirectional influences on risk, suggesting potential protective effects with optimization of factors such as estimated blood loss, surgical duration, and comorbidity burden. CONCLUSIONS:Meticulous management of patients with high comorbidity burdens or Medicaid insurance who are admitted non-electively or spend clinically indicated time in the intensive care unit (ICU) during their hospitalization course may be warranted to reduce their risk of unanticipated prolonged LOS following spine surgery. LEVEL OF EVIDENCE/METHODS:Prognostic Level III. See Instructions for Authors for a complete description of levels of evidence.

BACKGROUND:/Objective: With a growing aging population in the United States, the number of operative lumbar spine pathologies continues to grow. Therefore, our objective was to estimate the future demand for lumbar spine surgery volumes for the United States to the year 2040. METHODS:The National/Nationwide Inpatient Sample was queried for years 2003-2015 for anterior interbody and posterior lumbar fusions (ALIF, PLF) to create national estimates of procedural volumes for those years. The average age and comorbidity burden was characterized, and Poisson modeling controlling for age and sex allowed for surgical volume prediction to 2040 in ten year increments. Age was grouped into categories (<25 years old, 25-34, 35-44, 45-54, 55-64, 65-74, 75-84, and >85), and estimates of surgical volumes for each age subgroup were created RESULTS: ALIF volume is expected to increase from 46,903 to 55,528, while PLF volume is expected to increase from 248,416 to 297,994 from 2020 to 2040. For ALIF, the largest increases are expected in the 45-54 year old (10,316 to 12,216) and 75-84 year old (2,898 to 5,340) age groups. Similarly the largest increases in PLF will be seen in the 65-74 year old (71,087 to 77,786) and 75-84 year old (28,253 to 52,062) age groups. CONCLUSION/CONCLUSIONS:The large increases in expected volumes of ALIF and PLF could necessitate training of more spinal surgeons and an examination of projected costs. Further analyses are needed to characterize the needs of this increasingly large population of surgical patients.

OBJECTIVE:Few studies have examined the impact of teaching status and location on outcomes in subarachnoid hemorrhage (SAH). The objective of the present study was to compare mortality and functional outcomes between urban teaching, urban nonteaching, and rural centers for hospitalizations with SAH. METHODS:The National Inpatient Sample (NIS) years 2003-2016 was queried for hospitalizations with aneurysmal SAH from 2003 to 2017. Cohorts treated at urban teaching, urban nonteaching, and rural centers were compared with the urban teaching center cohort acting as the reference. The NIS-SOM, a validated measure of SAH functional outcome, was used as a co-primary outcome with mortality. Multivariable models adjusted for age, sex, NIH-SSS score, hypertension, and hospital bedsize. Trends in SAH mortality rates were calculated. RESULTS:There were 379,716 SAH hospitalizations at urban teaching centers, 105,638 at urban nonteaching centers, and 17,165 at rural centers. Adjusted mortality rates for urban teaching centers were lower than urban nonteaching (21.90% vs. 25.00%, p<0.0001) and rural (21.90% vs. 30.90%, p<0.0001) centers. While urban teaching (24.74% to 21.22%) and urban nonteaching (24.78% to 23.68%) had decreases in mortality rates over the study period, rural hospitals showed increased mortality rates (25.67% to 33.38%). CONCLUSIONS:Rural and urban nonteaching centers have higher rates of mortality from SAH than urban teaching centers. Further study is necessary to understand drivers of these differences.

BACKGROUND:Unplanned hospital readmissions constitute a significant cost burden in healthcare. Identifying factors contributing to readmission risk presents opportunities for actionable change to reduce readmission rates. OBJECTIVE:To combine machine learning classification and feature importance analysis to identify drivers of readmission in a large cohort of spine patients. METHODS:Cases involving surgical procedures for degenerative spine conditions between 2008 and 2016 were retrospectively reviewed. Of 11 150 cases, 396 patients (3.6%) experienced an unplanned hospital readmission within 30 d of discharge. Over 75 pre-discharge variables were collected and categorized into demographic, perioperative, and resource utilization feature domains. Random forest classification was used to construct predictive models for readmission from feature domains. An ensemble tree-specific method was used to quantify and rank features by relative importance. RESULTS:In the demographics domain, age and comorbidity burden were the most important features for readmission prediction. Surgical duration and intraoperative oral morphine equivalents were the most important perioperative features, whereas total direct cost and length of stay were most important in the resource utilization domain. In supervised learning experiments for predicting readmission, the demographic domain model performed the best alone, suggesting that demographic features may contribute more to readmission risk than perioperative variables following spine surgery. A predictive model, created using only enriched features showing substantial importance, demonstrated improved predictive capacity compared to previous models, and approached the performance of state-of-the-art, deep-learning models for readmission. CONCLUSION/CONCLUSIONS:This strategy provides insight into global patterns of feature importance and better understanding of drivers of readmissions following spine surgery.