Incorporating temporal EHR data in predictive models for risk stratification of renal function deterioration.

Predictive models built using temporal data in electronic health records (EHRs) can potentially play a major role in improving management of chronic diseases. However, these data present a multitude of technical challenges, including irregular sampling of data and varying length of available patient history. In this paper, we describe and evaluate three different approaches that use machine learning to build predictive models using temporal EHR data of a patient. The first approach is a commonly used non-temporal approach that aggregates values of the predictors in the patient's medical history. The other two approaches exploit the temporal dynamics of the data. The two temporal approaches vary in how they model temporal information and handle missing data. Using data from the EHR of Mount Sinai Medical Center, we learned and evaluated the models in the context of predicting loss of estimated glomerular filtration rate (eGFR), the most common assessment of kidney function. Our results show that incorporating temporal information in patient's medical history can lead to better prediction of loss of kidney function. They also demonstrate that exactly how this information is incorporated is important. In particular, our results demonstrate that the relative importance of different predictors varies over time, and that using multi-task learning to account for this is an appropriate way to robustly capture the temporal dynamics in EHR data. Using a case study, we also demonstrate how the multi-task learning based model can yield predictive models with better performance for identifying patients at high risk of short-term loss of kidney function.

Acoustic Voice and Speech Biomarkers of Treatment Status during Hospitalization for Acute Decompensated Heart Failure.

This study investigates acoustic voice and speech features as biomarkers for acute decompensated heart failure (ADHF), a serious escalation of heart failure symptoms including breathlessness and fatigue. ADHF-related systemic fluid accumulation in the lungs and laryngeal tissues is hypothesized to affect phonation and respiration for speech. A set of daily spoken recordings from 52 patients undergoing inpatient ADHF treatment was analyzed to identify voice and speech biomarkers for ADHF and to examine the trajectory of biomarkers during treatment. Results indicated that speakers produce more stable phonation, a more creaky voice, faster speech rates, and longer phrases after ADHF treatment compared to their pre-treatment voices. This project builds on work to develop a method of monitoring ADHF using speech biomarkers and presents a more detailed understanding of relevant voice and speech features.

Predicting outcomes in patients with aortic stenosis using machine learning: the Aortic Stenosis Risk (ASteRisk) score.

OBJECTIVE: To use echocardiographic and clinical features to develop an explainable clinical risk prediction model in patients with aortic stenosis (AS), including those with low-gradient AS (LGAS), using machine learning (ML). METHODS: In 1130 patients with moderate or severe AS, we used bootstrap lasso regression (BLR), an ML method, to identify echocardiographic and clinical features important for predicting the combined outcome of all-cause mortality or aortic valve replacement (AVR) within 5 years after the initial echocardiogram. A separate hold out set, from a different centre (n=540), was used to test the generality of the model. We also evaluated model performance with respect to each outcome separately and in different subgroups, including patients with LGAS. RESULTS: Out of 69 available variables, 26 features were identified as predictive by BLR and expert knowledge was used to further reduce this set to 9 easily available and input features without loss of efficacy. A ridge logistic regression model constructed using these features had an area under the receiver operating characteristic curve (AUC) of 0.74 for the combined outcome of mortality/AVR. The model reliably identified patients at high risk of death in years 2-5 (HRs ≥2.0, upper vs other quartiles, for years 2-5, p<0.05, p=not significant in year 1) and was also predictive in the cohort with LGAS (n=383, HRs≥3.3, p<0.05). The model performed similarly well in the independent hold out set (AUC 0.78, HR ≥2.5 in years 1-5, p<0.05). CONCLUSION: In two separate longitudinal databases, ML identified prognostic features and produced an algorithm that predicts outcome for up to 5 years of follow-up in patients with AS, including patients with LGAS. Our algorithm, the Aortic Stenosis Risk (ASteRisk) score, is available online for public use.

Do as AI say: susceptibility in deployment of clinical decision-aids.

Artificial intelligence (AI) models for decision support have been developed for clinical settings such as radiology, but little work evaluates the potential impact of such systems. In this study, physicians received chest X-rays and diagnostic advice, some of which was inaccurate, and were asked to evaluate advice quality and make diagnoses. All advice was generated by human experts, but some was labeled as coming from an AI system. As a group, radiologists rated advice as lower quality when it appeared to come from an AI system; physicians with less task-expertise did not. Diagnostic accuracy was significantly worse when participants received inaccurate advice, regardless of the purported source. This work raises important considerations for how advice, AI and non-AI, should be deployed in clinical environments.

Learning from Few Subjects with Large Amounts of Voice Monitoring Data.

Recently, researchers have started training high complexity machine learning models to clinical tasks, often improving upon previous benchmarks. However, more often than not, these methods require large amounts of supervision to provide good generalization guarantees. When applied to data coming from small cohorts and long monitoring periods these models are prone to overfit to subject-identifying features. Since obtaining large amounts of labels is usually not practical in many scenarios, expert-driven knowledge of the task is a common technique to prevent overfitting. We present a two-step learning approach that is able to generalize under these circumstances when applied to a voice monitoring dataset. Our approach decouples the feature learning stage and performs it in an unsupervised manner, removing the need for laborious feature engineering. We show the effectiveness of our proposed model on two voice monitoring related tasks. We evaluate the extracted features for classifying between patients with vocal fold nodules and controls. We also demonstrate that the features capture pathology relevant information by showing that models trained on them are more accurate predicting vocal use for patients than for controls. Our proposed method is able to generalize to unseen subjects and across learning tasks while matching state-of-the-art results.

SMART--an integrated wireless system for monitoring unattended patients.

Monitoring vital signs and locations of certain classes of ambulatory patients can be useful in overcrowded emergency departments and at disaster scenes, both on-site and during transportation. To be useful, such monitoring needs to be portable and low cost, and have minimal adverse impact on emergency personnel, e.g., by not raising an excessive number of alarms. The SMART (Scalable Medical Alert Response Technology) system integrates wireless patient monitoring (ECG, SpO(2)), geo-positioning, signal processing, targeted alerting, and a wireless interface for caregivers. A prototype implementation of SMART was piloted in the waiting area of an emergency department and evaluated with 145 post-triage patients. System deployment aspects were also evaluated during a small-scale disaster-drill exercise.

Ambulatory assessment of phonotraumatic vocal hyperfunction using glottal airflow measures estimated from neck-surface acceleration.

Phonotraumatic vocal hyperfunction (PVH) is associated with chronic misuse and/or abuse of voice that can result in lesions such as vocal fold nodules. The clinical aerodynamic assessment of vocal function has been recently shown to differentiate between patients with PVH and healthy controls to provide meaningful insight into pathophysiological mechanisms associated with these disorders. However, all current clinical assessment of PVH is incomplete because of its inability to objectively identify the type and extent of detrimental phonatory function that is associated with PVH during daily voice use. The current study sought to address this issue by incorporating, for the first time in a comprehensive ambulatory assessment, glottal airflow parameters estimated from a neck-mounted accelerometer and recorded to a smartphone-based voice monitor. We tested this approach on 48 patients with vocal fold nodules and 48 matched healthy-control subjects who each wore the voice monitor for a week. Seven glottal airflow features were estimated every 50 ms using an impedance-based inverse filtering scheme, and seven high-order summary statistics of each feature were computed every 5 minutes over voiced segments. Based on a univariate hypothesis testing, eight glottal airflow summary statistics were found to be statistically different between patient and healthy-control groups. L1-regularized logistic regression for a supervised classification task yielded a mean (standard deviation) area under the ROC curve of 0.82 (0.25) and an accuracy of 0.83 (0.14). These results outperform the state-of-the-art classification for the same classification task and provide a new avenue to improve the assessment and treatment of hyperfunctional voice disorders.

Beatquency domain and machine learning improve prediction of cardiovascular death after acute coronary syndrome.

Frequency domain measures of heart rate variability (HRV) are associated with adverse events after a myocardial infarction. However, patterns in the traditional frequency domain (measured in Hz, or cycles per second) may capture different cardiac phenomena at different heart rates. An alternative is to consider frequency with respect to heartbeats, or beatquency. We compared the use of frequency and beatquency domains to predict patient risk after an acute coronary syndrome. We then determined whether machine learning could further improve the predictive performance. We first evaluated the use of pre-defined frequency and beatquency bands in a clinical trial dataset (N = 2302) for the HRV risk measure LF/HF (the ratio of low frequency to high frequency power). Relative to frequency, beatquency improved the ability of LF/HF to predict cardiovascular death within one year (Area Under the Curve, or AUC, of 0.730 vs. 0.704, p < 0.001). Next, we used machine learning to learn frequency and beatquency bands with optimal predictive power, which further improved the AUC for beatquency to 0.753 (p < 0.001), but not for frequency. Results in additional validation datasets (N = 2255 and N = 765) were similar. Our results suggest that beatquency and machine learning provide valuable tools in physiological studies of HRV.

Uncovering Voice Misuse Using Symbolic Mismatch.

Voice disorders affect an estimated 14 million working-aged Americans, and many more worldwide. We present the first large scale study of vocal misuse based on long-term ambulatory data collected by an accelerometer placed on the neck. We investigate an unsupervised data mining approach to uncovering latent information about voice misuse. We segment signals from over 253 days of data from 22 subjects into over a hundred million single glottal pulses (closures of the vocal folds), cluster segments into symbols, and use symbolic mismatch to uncover differences between patients and matched controls, and between patients pre- and post-treatment. Our results show significant behavioral differences between patients and controls, as well as between some pre- and post-treatment patients. Our proposed approach provides an objective basis for helping diagnose behavioral voice disorders, and is a first step towards a more data-driven understanding of the impact of voice therapy.

Using Ambulatory Voice Monitoring to Investigate Common Voice Disorders: Research Update.

Many common voice disorders are chronic or recurring conditions that are likely to result from inefficient and/or abusive patterns of vocal behavior, referred to as vocal hyperfunction. The clinical management of hyperfunctional voice disorders would be greatly enhanced by the ability to monitor and quantify detrimental vocal behaviors during an individual's activities of daily life. This paper provides an update on ongoing work that uses a miniature accelerometer on the neck surface below the larynx to collect a large set of ambulatory data on patients with hyperfunctional voice disorders (before and after treatment) and matched-control subjects. Three types of analysis approaches are being employed in an effort to identify the best set of measures for differentiating among hyperfunctional and normal patterns of vocal behavior: (1) ambulatory measures of voice use that include vocal dose and voice quality correlates, (2) aerodynamic measures based on glottal airflow estimates extracted from the accelerometer signal using subject-specific vocal system models, and (3) classification based on machine learning and pattern recognition approaches that have been used successfully in analyzing long-term recordings of other physiological signals. Preliminary results demonstrate the potential for ambulatory voice monitoring to improve the diagnosis and treatment of common hyperfunctional voice disorders.

ECG morphological variability in beat space for risk stratification after acute coronary syndrome.

BACKGROUND: Identification of patients who are at high risk of adverse cardiovascular events after an acute coronary syndrome (ACS) remains a major challenge in clinical cardiology. We hypothesized that quantifying variability in electrocardiogram (ECG) morphology may improve risk stratification post-ACS. METHODS AND RESULTS: We developed a new metric to quantify beat-to-beat morphologic changes in the ECG: morphologic variability in beat space (MVB), and compared our metric to published ECG metrics (heart rate variability [HRV], deceleration capacity [DC], T-wave alternans, heart rate turbulence, and severe autonomic failure). We tested the ability of these metrics to identify patients at high risk of cardiovascular death (CVD) using 1082 patients (1-year CVD rate, 4.5%) from the MERLIN-TIMI 36 (Metabolic Efficiency with Ranolazine for Less Ischemia in Non-ST-Elevation Acute Coronary Syndrome-Thrombolysis in Myocardial Infarction 36) clinical trial. DC, HRV/low frequency-high frequency, and MVB were all associated with CVD (hazard ratios [HRs] from 2.1 to 2.3 [P<0.05 for all] after adjusting for the TIMI risk score [TRS], left ventricular ejection fraction [LVEF], and B-type natriuretic peptide [BNP]). In a cohort with low-to-moderate TRS (N=864; 1-year CVD rate, 2.7%), only MVB was significantly associated with CVD (HR, 3.0; P=0.01, after adjusting for LVEF and BNP). CONCLUSIONS: ECG morphological variability in beat space contains prognostic information complementary to the clinical variables, LVEF and BNP, in patients with low-to-moderate TRS. ECG metrics could help to risk stratify patients who might not otherwise be considered at high risk of CVD post-ACS.

Learning Data-Driven Patient Risk Stratification Models for Clostridium difficile.

BACKGROUND: Although many risk factors are well known, Clostridium difficile infection (CDI) continues to be a significant problem throughout the world. The purpose of this study was to develop and validate a data-driven, hospital-specific risk stratification procedure for estimating the probability that an inpatient will test positive for C difficile. METHODS: We consider electronic medical record (EMR) data from patients admitted for ≥24 hours to a large urban hospital in the U.S. between April 2011 and April 2013. Predictive models were constructed using L2-regularized logistic regression and data from the first year. The number of observational variables considered varied from a small set of well known risk factors readily available to a physician to over 10 000 variables automatically extracted from the EMR. Each model was evaluated on holdout admission data from the following year. A total of 34 846 admissions with 372 cases of CDI was used to train the model. RESULTS: Applied to the separate validation set of 34 722 admissions with 355 cases of CDI, the model that made use of the additional EMR data yielded an area under the receiver operating characteristic curve (AUROC) of 0.81 (95% confidence interval [CI], .79-.83), and it significantly outperformed the model that considered only the small set of known clinical risk factors, AUROC of 0.71 (95% CI, .69-.75). CONCLUSIONS: Automated risk stratification of patients based on the contents of their EMRs can be used to accurately identify a high-risk population of patients. The proposed method holds promise for enabling the selective allocation of interventions aimed at reducing the rate of CDI.

Learning to detect vocal hyperfunction from ambulatory neck-surface acceleration features: initial results for vocal fold nodules.

Voice disorders are medical conditions that often result from vocal abuse/misuse which is referred to generically as vocal hyperfunction. Standard voice assessment approaches cannot accurately determine the actual nature, prevalence, and pathological impact of hyperfunctional vocal behaviors because such behaviors can vary greatly across the course of an individual's typical day and may not be clearly demonstrated during a brief clinical encounter. Thus, it would be clinically valuable to develop noninvasive ambulatory measures that can reliably differentiate vocal hyperfunction from normal patterns of vocal behavior. As an initial step toward this goal we used an accelerometer taped to the neck surface to provide a continuous, noninvasive acceleration signal designed to capture some aspects of vocal behavior related to vocal cord nodules, a common manifestation of vocal hyperfunction. We gathered data from 12 female adult patients diagnosed with vocal fold nodules and 12 control speakers matched for age and occupation. We derived features from weeklong neck-surface acceleration recordings by using distributions of sound pressure level and fundamental frequency over 5-min windows of the acceleration signal and normalized these features so that intersubject comparisons were meaningful. We then used supervised machine learning to show that the two groups exhibit distinct vocal behaviors that can be detected using the acceleration signal. We were able to correctly classify 22 of the 24 subjects, suggesting that in the future measures of the acceleration signal could be used to detect patients with the types of aberrant vocal behaviors that are associated with hyperfunctional voice disorders.

A comparison of non-symmetric entropy-based classification trees and support vector machine for cardiovascular risk stratification.

Classification tree-based risk stratification models generate easily interpretable classification rules. This feature makes classification tree-based models appealing for use in a clinical setting, provided that they have comparable accuracy to other methods. In this paper, we present and evaluate the performance of a non-symmetric entropy-based classification tree algorithm. The algorithm is designed to accommodate class imbalance found in many medical datasets. We evaluate the performance of this algorithm, and compare it to that of SVM-based classifiers, when applied to 4219 non-ST elevation acute coronary syndrome patients. We generated SVM-based classifiers using three different strategies for handling class imbalance: cost-sensitive SVM learning, synthetic minority oversampling (SMOTE), and random majority undersampling. We used both linear and radial basis kernel-based SVMs. Our classification tree models outperformed SVM-based classifiers generated using each of the three techniques. On average, the classification tree models yielded a 14% improvement in G-score and a 21% improvement in F-score relative to the linear SVM classifiers with the best performance. Similarly, our classification tree models yielded a 12% improvement in G-score and a 21% improvement in the F-score over the best RBF kernel-based SVM classifiers.

Patient-specific ventricular beat classification without patient-specific expert knowledge: a transfer learning approach.

We present an adaptive binary classification algorithm, based on transductive transfer learning. We illustrate the method in the context of electrocardiogram (ECG) analysis. Knowledge gained from a population of patients is automatically adapted to patients' records to accurately detect ectopic beats. On patients from the MIT-BIH Arrhythmia Database, we achieve a median sensitivity of 94.59% and positive predictive value of 96.24%, for the binary classification task of separating premature ventricular contractions (PVCs), a type of ectopic beat, from non-PVCs.

Computationally generated cardiac biomarkers for risk stratification after acute coronary syndrome.

The existing tools for estimating the risk of death in patients after they experience acute coronary syndrome are commonly based on echocardiography and clinical risk scores (for example, the TIMI risk score). These identify a small group of high-risk patients who account for only a minority of the deaths that occur in patients after acute coronary syndrome. Here, we investigated the use of three computationally generated cardiac biomarkers for risk stratification in this population: morphologic variability (MV), symbolic mismatch (SM), and heart rate motifs (HRM). We derived these biomarkers from time-series analyses of continuous electrocardiographic data collected from patients in the TIMI-DISPERSE2 clinical trial through machine learning and data mining methods designed to extract information that is difficult to visualize directly in these data. We evaluated these biomarkers in a blinded, prespecified, and fully automated study on more than 4500 patients in the MERLIN-TIMI36 (Metabolic Efficiency with Ranolazine for Less Ischemia in Non-ST-Elevation Acute Coronary Syndrome-Thrombolysis in Myocardial Infarction 36) clinical trial. Our results showed a strong association between all three computationally generated cardiac biomarkers and cardiovascular death in the MERLIN-TIMI36 trial over a 2-year period after acute coronary syndrome. Moreover, the information in each of these biomarkers was independent of the information in the others and independent of the information provided by existing clinical risk scores, electrocardiographic metrics, and echocardiography. The addition of MV, SM, and HRM to existing metrics significantly improved model discrimination, as well as the precision and recall of prediction rules based on left ventricular ejection fraction. These biomarkers can be extracted from data that are routinely captured from patients with acute coronary syndrome and will allow for more accurate risk stratification and potentially for better patient treatment.

Spectral energy of ECG morphologic differences to predict death.

Unstable conduction system bifurcations following ischemia and infarction are associated with variations in the electrocardiographic activity spanning the heart beat. In this paper, we investigate a spectral energy measure of morphologic differences (SE-MD) that quantifies aspects of these changes. Our measure uses a dynamic time-warping approach to compute the time-aligned morphology differences between pairs of successive sinus beats in an electrocardiographic signal. While comparing beats, the entire heart beat signal is analyzed in order to capture changes affecting both depolarization and repolarization. We show that variations in electrocardiographic activity associated with death can be distinguished by their spectral characteristics. We developed the SE-MD metric on holter data from 764 patients from the TIMI DISPERSE2 dataset and tested it on 600 patients from the TIMI MERLIN dataset. In the test population, high SE-MD was strongly associated with death over a 90 day period following non-ST-elevation acute coronary syndrome (HR 10.45, p < 0.001) and showed significant discriminative ability (c-statistic 0.85). In comparison with heart rate variability and deceleration capacity, SE-MD was also the most significant predictor of death in the study population. Furthermore, SE-MD had low correlation with these other measures, suggesting that complementary use of the risk variables may allow for more complete assessment of cardiac health.

Relation of death within 90 days of non-ST-elevation acute coronary syndromes to variability in electrocardiographic morphology.

Electrocardiographic measures can facilitate the identification of patients at risk of death after acute coronary syndromes. This study evaluates a new risk metric, morphologic variability (MV), which measures beat-to-beat variability in the shape of the entire heart beat signal. This metric is analogous to heart rate variability (HRV) approaches, which focus on beat-to-beat changes in the heart rate. MV was calculated using a dynamic time-warping technique in 764 patients from the DISPERSE2 (TIMI 33) trial for whom 24-hour continuous electrocardiograph was recorded within 48 hours of non-ST-elevation acute coronary syndrome. The patients were evaluated during a 90-day follow-up for the end point of death. Patients with high MV showed an increased risk of death during follow-up (hazard ratio 8.46; p <0.001). The relationship between high MV and death could be observed even after adjusting for baseline clinical characteristics and HRV measures (adjusted hazard ratio 6.91; p = 0.001). Moreover, the correlation between MV and HRV was low (R < or =0.25). These findings were consistent among several subgroups, including patients under the age of 65 and those with no history of diabetes or hyperlipidemia. In conclusion, our results suggest that increased variation in the entire heart beat morphology is associated with a considerably elevated risk of death and may provide information complementary to the analysis of heart rate.

Reducing the number of channels for an ambulatory patient-specific EEG-based epileptic seizure detector by applying recursive feature elimination.

We are building an ambulatory version of a patient-specific epileptic seizure detector based on scalp EEG signals. Since patients have to wear the electrodes all the time, it is desirable to use the minimum number of electrodes needed to achieve good performance. In this paper, we describe a method that uses recursive feature elimination (RFE) to design detectors that use small numbers of electrodes. We also present results that indicate that the appropriate number of electrodes varies across patients. It is frequently the case that a surprisingly small number of electrodes, sometimes as few as two, suffices to construct a detector with expected performance comparable to that of detectors that use a full twenty-one-channel montage.