Daniel Fabbri

A Scalable Framework to Detect Personal Health Mentions on Twitter

Background Biomedical research has traditionally been conducted via surveys and the analysis of medical records. However, these resources are limited in their content, such that non-traditional domains (eg, online forums and social media) have an opportunity to supplement the view of an individual’s health. Objective The objective of this study was to develop a scalable framework to detect personal health status mentions on Twitter and assess the extent to which such information is disclosed. Methods We collected more than 250 million tweets via the Twitter streaming API over a 2-month period in 2014. The corpus was filtered down to approximately 250,000 tweets, stratified across 34 high-impact health issues, based on guidance from the Medical Expenditure Panel Survey. We created a labeled corpus of several thousand tweets via a survey, administered over Amazon Mechanical Turk, that documents when terms correspond to mentions of personal health issues or an alternative (eg, a metaphor). We engineered a scalable classifier for personal health mentions via feature selection and assessed its potential over the health issues. We further investigated the utility of the tweets by determining the extent to which Twitter users disclose personal health status. Results Our investigation yielded several notable findings. First, we find that tweets from a small subset of the health issues can train a scalable classifier to detect health mentions. Specifically, training on 2000 tweets from four health issues (cancer, depression, hypertension, and leukemia) yielded a classifier with precision of 0.77 on all 34 health issues. Second, Twitter users disclosed personal health status for all health issues. Notably, personal health status was disclosed over 50% of the time for 11 out of 34 (33%) investigated health issues. Third, the disclosure rate was dependent on the health issue in a statistically significant manner (P<.001). For instance, more than 80% of the tweets about migraines (83/100) and allergies (85/100) communicated personal health status, while only around 10% of the tweets about obesity (13/100) and heart attack (12/100) did so. Fourth, the likelihood that people disclose their own versus other people’s health status was dependent on health issue in a statistically significant manner as well (P<.001). For example, 69% (69/100) of the insomnia tweets disclosed the author’s status, while only 1% (1/100) disclosed another person’s status. By contrast, 1% (1/100) of the Down syndrome tweets disclosed the author’s status, while 21% (21/100) disclosed another person’s status. Conclusions It is possible to automatically detect personal health status mentions on Twitter in a scalable manner. These mentions correspond to the health issues of the Twitter users themselves, but also other individuals. Though this study did not investigate the veracity of such statements, we anticipate such information may be useful in supplementing traditional health-related sources for research purposes.

Complexity of medical decision-making in care provided by surgeons through patient portals

BACKGROUND Patient portals are online applications that allow patients to interact with healthcare organizations and view information. Portal messages exchanged between patients and providers contain diverse types of communications, including delivery of medical care. The types of communications and complexity of medical decision-making in portal messages sent to surgeons have not been studied. MATERIALS AND METHODS We obtained all message threads initiated by patients and exchanged with surgical providers through the Vanderbilt University Medical Center patient portal from June 1 to December 31, 2014. Five hundred randomly selected messages were manually analyzed by two research team members to determine the types of communication (i.e., informational, medical, logistical, or social), whether medical care was delivered, and complexity of medical decision-making as defined for outpatient billing in each message thread. RESULTS A total of 9408 message threads were sent to 401 surgical providers during the study period. In the 500 threads selected for detailed analysis, 1293 distinct issues were communicated, with an average of 2.6 issues per thread. Medical needs were communicated in 453 message threads (90.6%). Further, 339 message threads (67.8%) contained medical decision-making. Overall complexity of medical decision-making was straightforward in 210 messages (62%), low in 102 messages (30%), and moderate in 27 messages (8%). No highly complex decisions were made over portal messaging. CONCLUSIONS Through patient portal messages, surgeons deliver substantial medical care with varied levels of medical complexity. Models for compensation of online care must be developed as consumer and surgeon adoption of these technologies increases.

Toward content-based image retrieval with deep convolutional neural networks

Content-based image retrieval (CBIR) offers the potential to identify similar case histories, understand rare disorders, and eventually, improve patient care. Recent advances in database capacity, algorithm efficiency, and deep Convolutional Neural Networks (dCNN), a machine learning technique, have enabled great CBIR success for general photographic images. Here, we investigate applying the leading ImageNet CBIR technique to clinically acquired medical images captured by the Vanderbilt Medical Center. Briefly, we (1) constructed a dCNN with four hidden layers, reducing dimensionality of an input scaled to 128x128 to an output encoded layer of 4x384, (2) trained the network using back-propagation 1 million random magnetic resonance (MR) and computed tomography (CT) images, (3) labeled an independent set of 2100 images, and (4) evaluated classifiers on the projection of the labeled images into manifold space. Quantitative results were disappointing (averaging a true positive rate of only 20%); however, the data suggest that improvements would be possible with more evenly distributed sampling across labels and potential re-grouping of label structures. This preliminary effort at automated classification of medical images with ImageNet is promising, but shows that more work is needed beyond direct adaptation of existing techniques.

Predicting Discharge Dates From the NICU Using Progress Note Data

BACKGROUND AND OBJECTIVES: Discharging patients from the NICU may be delayed for nonmedical reasons including the need for medical equipment, parental education, and children’s services. We describe a method to predict which patients will be medically ready for discharge in the next 2 to 10 days, providing lead time to address nonmedical reasons for delayed discharge. METHODS: A retrospective study examined 26 features (17 extracted, 9 engineered) from daily progress notes of 4693 patients (103 206 patient-days) from the NICU of a large, academic children’s hospital. These data were used to develop a supervised machine learning problem to predict days to discharge (DTD). Random forest classifiers were trained by using examined features and International Classification of Diseases, Ninth Revision–based subpopulations to determine the most important features. RESULTS: Three of the 4 subpopulations (premature, cardiac, gastrointestinal surgery) and all patients combined performed similarly at 2, 4, 7, and 10 DTD with area under the curve (AUC) ranging from 0.854 to 0.865 at 2 DTD and 0.723 to 0.729 at 10 DTD. Patients undergoing neurosurgery performed worse at every DTD measure, scoring 0.749 at 2 DTD and 0.614 at 10 DTD. This model was also able to identify important features and provide “rule-of-thumb” criteria for patients close to discharge. By using DTD equal to 4 and 2 features (oral percentage of feedings and weight), we constructed a model with an AUC of 0.843. CONCLUSIONS: Using clinical features from daily progress notes provides an accurate method to predict when patients in the NICU are nearing discharge.

Patient Stratification Using Electronic Health Records from a Chronic Disease Management Program

OBJECTIVE The goal of this study is to devise a machine learning framework to assist care coordination programs in prognostic stratification to design and deliver personalized care plans and to allocate financial and medical resources effectively. MATERIALS AND METHODS This study is based on a de-identified cohort of 2,521 hypertension patients from a chronic care coordination program at the Vanderbilt University Medical Center. Patients were modeled as vectors of features derived from electronic health records (EHRs) over a six-year period. We applied a stepwise regression to identify risk factors associated with a decrease in mean arterial pressure of at least 2 mmHg after program enrollment. The resulting features were subsequently validated via a logistic regression classifier. Finally, risk factors were applied to group the patients through model-based clustering. RESULTS We identified a set of predictive features that consisted of a mix of demographic, medication, and diagnostic concepts. Logistic regression over these features yielded an area under the ROC curve (AUC) of 0.71 (95% CI: [0.67, 0.76]). Based on these features, four clinically meaningful groups are identified through clustering - two of which represented patients with more severe disease profiles, while the remaining represented patients with mild disease profiles. DISCUSSION Patients with hypertension can exhibit significant variation in their blood pressure control status and responsiveness to therapy. Yet this work shows that a clustering analysis can generate more homogeneous patient groups, which may aid clinicians in designing and implementing customized care programs. CONCLUSION The study shows that predictive modeling and clustering using EHR data can be beneficial for providing a systematic, generalized approach for care providers to tailor their management approach based upon patient-level factors.

Structural Functional Associations of the Orbit in Thyroid Eye Disease: Kalman Filters to Track Extraocular Rectal Muscles

Pathologies of the optic nerve and orbit impact millions of Americans and quantitative assessment of the orbital structures on 3-D imaging would provide objective markers to enhance diagnostic accuracy, improve timely intervention, and eventually preserve visual function. Recent studies have shown that the multi-atlas methodology is suitable for identifying orbital structures, but challenges arise in the identification of the individual extraocular rectus muscles that control eye movement. This is increasingly problematic in diseased eyes, where these muscles often appear to fuse at the back of the orbit (at the resolution of clinical computed tomography imaging) due to inflammation or crowding. We propose the use of Kalman filters to track the muscles in three-dimensions to refine multi-atlas segmentation and resolve ambiguity due to imaging resolution, noise, and artifacts. The purpose of our study is to investigate a method of automatically generating orbital metrics from CT imaging and demonstrate the utility of the approach by correlating structural metrics of the eye orbit with clinical data and visual function measures in subjects with thyroid eye disease. The pilot study demonstrates that automatically calculated orbital metrics are strongly correlated with several clinical characteristics. Moreover, it is shown that the superior, inferior, medial and lateral rectus muscles obtained using Kalman filters are each correlated with different categories of functional deficit. These findings serve as foundation for further investigation in the use of CT imaging in the study, analysis and diagnosis of ocular diseases, specifically thyroid eye disease.

Hands Free Automatic Clinical Care Documentation: Opportunities for Motion Sensors and Cameras

Electronic care documentation continues to represent a significant time expense in many clinical environments. The fundamental challenge centers around the need to either (i) pause from the provision of care to document clinical activities, or (ii) document care after the encounter. This serial workflow prevents simultaneous care and documentation, which can result in extensive physician frustration, errors in documentation and clinical errors. As a result, medical scribes are increasingly being used in high throughput clinical areas to reduce the documentation burden on clinicians. However, the addition of personnel to create documentation is resource intensive. Hands free clinical documentation systems can alleviate these documentation problems [1–3]. Systems that rely on voice recognition have shown some promise [4], but suffer from performance deterioration in settings where (1) there is any significant degree of ambient noise or (2) multiple caregivers are present, (3) most acute care environments where voice commands are indistinguishable from conversations. A new approach to hands free automatic clinical care documentation is therefore needed to improve the functioning of our medical systems. Given the advances in off-the shelf sensors, such as wristworn motion sensors and cameras, it is possible to automatically generate a clinical care record by detecting the signatures associated with key clinical tasks [5–10]. Wrist-worn sensors contain accelerometers and gyroscopes, which can track motion and identify clinical interventions such as CPR [11, 12]. Cameras can be used to track physician hand position and their relationship with the patient’s body to construct an injury heatmap. Additionally, the fusion of motion sensors and cameras allows for even further opportunities for completing the clinical care record. Realistically, the addition of motion sensors and cameras will not be able to identify every aspect of clinical care in the near term. However, specific environments and interventions may be amenable to these methods, one of which is patient transport. During patient transport, patient movement is restricted and the set of possible interventions is limited. Moreover, receiving physicians can benefit from real-time and accurate documentation. If successful, hands free documentation can enable clinicians to focus on the clinical tasks at hand (i.e. the patient in front of them), effectively hand-off patients between providers, and allow receiving physicians to better prepare for patient arrivals. The medical community must leverage emerging technologies to improve clinical care. The automatic identification and documentation of key clinical concepts is an important next step in optimizing clinician resource utilization.

Discovering de facto diagnosis specialties

In health care institutions, medical specialty information may be lacking or inaccurate. Diagnosis histories offer information on which medical specialties may exist in practice, regardless of whether they have official codes. We refer to such specialties that are predicted with high certainty by diagnosis histories de facto diagnosis specialties. We aim to discover de facto diagnosis specialties under a general discovery--evaluation framework. Specifically, we employ a semi-supervised learning model and an unsupervised learning method for discovery. We further employ four supervised learning models for evaluation. We use one year of diagnosis histories from a major medical center, which consists of two data sets: one is fine-grained and the other is general. The semi-supervised learning model discovers a specialty for Breast Cancer on the fine-grained data set; while the unsupervised learning method confirms this discovery and suggests another specialty for Obesity on the larger general data set. The evaluation results reinforce that these two specialties can be recognized accurately by supervised learning models in comparison with 12 common diagnosis specialties defined by the Health Care Provider Taxonomy Code Set.

Schedule-based metrics for the evaluation of clinic performance and potential recovery of cancelled appointments

BACKGROUND Assessment of outpatient clinic performance is important to optimize patient access. Metrics based on schedule data may assist with assessment of operational efficiency and recovering cancelled appointments. OBJECTIVES To define schedule-based characteristics of clinic operations and to evaluate potential for recovery of cancelled appointments. METHODS Sixty-seven weekly cardiology clinics from a single provider over 18 months at an academic medical center were analyzed. Parameters included clinic slots eligible to have patients scheduled (available), slots occupied by appointments (scheduled), and slots for which patients attended the associated visit (appeared). Rates of usage (scheduled/available), appearance (appeared/scheduled), and utilization (appeared/available=usage rate*appearance rate) were calculated. Surplus slots were defined as the difference between available slots and slots occupied by patients that appeared. Cancellation lag-time was defined as the interval between a cancellation and the appointment time. If a patient did not notify the clinic regarding a non-appearance, cancellation lag-time was set to zero. To quantify the impact of a change in clinic operations on efficiency, these metrics were used to evaluate a different cardiologists clinic before and after its physical location moved. RESULTS For approximately 900 patient visits, usage and appearance rates were∼80%, yielding a utilization rate of ∼2/3. On average, there were nearly 8 surplus slots per clinic. Approximately 30% of cancellation lag-times had positive values and nearly half of positive cancellation lag-times were >3h, indicating potential for recovery of those appointments. The intervention analysis showed that usage rate decreased and surplus slots per clinic increased significantly after a change in clinic location. CONCLUSIONS Schedule-based analysis provides a framework to assess changes to clinic operations, identify mechanisms underlying inefficiency, and suggest solutions for improving clinic performance (i.e. more schedulers in response to low usage rates). Cancellation lag-time analysis suggests recovering a portion of same-day cancellations is plausible.

Deep learning for brain tumor classification

Recent research has shown that deep learning methods have performed well on supervised machine learning, image classification tasks. The purpose of this study is to apply deep learning methods to classify brain images with different tumor types: meningioma, glioma, and pituitary. A dataset was publicly released containing 3,064 T1-weighted contrast enhanced MRI (CE-MRI) brain images from 233 patients with either meningioma, glioma, or pituitary tumors split across axial, coronal, or sagittal planes. This research focuses on the 989 axial images from 191 patients in order to avoid confusing the neural networks with three different planes containing the same diagnosis. Two types of neural networks were used in classification: fully connected and convolutional neural networks. Within these two categories, further tests were computed via the augmentation of the original 512×512 axial images. Training neural networks over the axial data has proven to be accurate in its classifications with an average five-fold cross validation of 91.43% on the best trained neural network. This result demonstrates that a more general method (i.e. deep learning) can outperform specialized methods that require image dilation and ring-forming subregions on tumors.