Alexander Van Esbroeck

Predicting readmission risk with institution-specific prediction models

The ability to predict patient readmission risk is extremely valuable for hospitals, especially under the Hospital Readmission Reduction Program (HRRP) of the Center for Medicare and Medicaid Services (CMS) which went into effect starting October 1, 2012. There is a plethora of work in the literature that deals with developing readmission risk prediction models, but most of them do not have sufficient prediction accuracy to be deployed in a clinical setting, partly because different hospitals may have different characteristics in their patient populations. In this work we experimented with a generic framework for institution-specific readmission risk prediction, which takes patient data from a single institution and produces a statistical risk prediction model optimized for that particular institution and optionally condition specific. This provides great flexibility in model building, and is also able to provide institution-specific insights in its readmitted patient population. We showcase some initial results at three institutions for Heart Failure (HF), Acute Myocardial Infarction (AMI) and Pneumonia (PN) patients. The developed models yield better prediction accuracy than the ones present in the literature.

Predicting Readmission Risk with Institution Specific Prediction Models

The ability to predict patient readmission risk is extremely valuable for hospitals, especially under the Hospital Readmission Reduction Program (HRRP) of the Center for Medicare and Medicaid Services (CMS) which went into effect starting October 1, 2012. There is a plethora of work in the literature that deals with developing readmission risk prediction models, but most of them do not have sufficient prediction accuracy to be deployed in a clinical setting, partly because different hospitals may have different characteristics in their patient populations. In this work we experimented with a generic framework for institution-specific readmission risk prediction, which takes patient data from a single institution and produces a statistical risk prediction model optimized for that particular institution and optionally condition specific. This provides great flexibility in model building, and is also able to provide institution-specific insights in its readmitted patient population. We showcase some initial results at three institutions for Heart Failure (HF), Acute Myocardial Infarction (AMI) and Pneumonia (PN) patients. The developed models yield better prediction accuracy than the ones present in the literature.

Quantifying surgical complexity with machine learning: Looking beyond patient factors to improve surgical models

OBJECTIVE To investigate the use of machine learning to empirically determine the risk of individual surgical procedures and to improve surgical models with this information. METHODS American College of Surgeons National Surgical Quality Improvement Program (ACS NSQIP) data from 2005 to 2009 were used to train support vector machine (SVM) classifiers to learn the relationship between textual constructs in current procedural terminology (CPT) descriptions and mortality, morbidity, Clavien 4 complications, and surgical-site infections (SSI) within 30 days of surgery. The procedural risk scores produced by the SVM classifiers were validated on data from 2010 in univariate and multivariate analyses. RESULTS The procedural risk scores produced by the SVM classifiers achieved moderate-to-high levels of discrimination in univariate analyses (area under receiver operating characteristic curve: 0.871 for mortality, 0.789 for morbidity, 0.791 for SSI, 0.845 for Clavien 4 complications). Addition of these scores also substantially improved multivariate models comprising patient factors and previously proposed correlates of procedural risk (net reclassification improvement and integrated discrimination improvement: 0.54 and 0.001 for mortality, 0.46 and 0.011 for morbidity, 0.68 and 0.022 for SSI, 0.44 and 0.001 for Clavien 4 complications; P < .05 for all comparisons). Similar improvements were noted in discrimination and calibration for other statistical measures, and in subcohorts comprising patients with general or vascular surgery. CONCLUSION Machine learning provides clinically useful estimates of surgical risk for individual procedures. This information can be measured in an entirely data-driven manner and substantially improves multifactorial models to predict postoperative complications.

Data-driven modeling of sleep states from EEG

Sleep analysis is critical for the diagnosis, treatment, and understanding of sleep disorders. However, the current standards for sleep analysis are widely considered oversimplified and problematic. The ability to automatically annotate different states during a night of sleep in a manner that is more descriptive than current standards, as well as the ability to train these models on a patient-by-patient basis, would provide a complementary approach for sleep analysis. We present a method that discovers latent structure in sleep EEG recordings, by extracting symbols from the continuous EEG signal and learning “topics” for a recording. These sleep topics are derived in a fully automatic and data-driven manner, and can represent the data with mixtures of states. The proposed method allows for identification of states in a patient-specific way, as opposed to the one-size-fits-all approach of the current standard. We demonstrate on a publicly available dataset of 15 sleep recordings that not only do the states discovered by this approach encompass the standard sleep stage structure, they provide additional information about sleep architecture with the potential to provide new insights into sleep disorders.

High throughput protease profiling comprehensively defines active site specificity for thrombin and ADAMTS13

We have combined random 6 amino acid substrate phage display with high throughput sequencing to comprehensively define the active site specificity of the serine protease thrombin and the metalloprotease ADAMTS13. The substrate motif for thrombin was determined by >6,700 cleaved peptides, and was highly concordant with previous studies. In contrast, ADAMTS13 cleaved only 96 peptides (out of >107 sequences), with no apparent consensus motif. However, when the hexapeptide library was substituted into the P3-P3′ interval of VWF73, an exosite-engaging substrate of ADAMTS13, 1670 unique peptides were cleaved. ADAMTS13 exhibited a general preference for aliphatic amino acids throughout the P3-P3′ interval, except at P2 where Arg was tolerated. The cleaved peptides assembled into a motif dominated by P3 Leu, and bulky aliphatic residues at P1 and P1′. Overall, the P3-P2′ amino acid sequence of von Willebrand Factor appears optimally evolved for ADAMTS13 recognition. These data confirm the critical role of exosite engagement for substrates to gain access to the active site of ADAMTS13, and define the substrate recognition motif for ADAMTS13. Combining substrate phage display with high throughput sequencing is a powerful approach for comprehensively defining the active site specificity of proteases.

Quality and outcomes reporting in trauma using International Statistical Classification for Diseases, Ninth Revision codes

BACKGROUND Use of the trauma and injury severity score (TRISS) for quality and outcomes assessment is challenged by the need for laborious collection of demographic and physiological data. We hypothesize that a novel stratification approach based on International Statistical Classification for Diseases, Ninth Revision (ICD-9) data that are readily available for trauma patients provides a more accurate and more easily obtainable alternative to TRISS with the potential for widespread use. METHODS Data from the ACS National Trauma Data Bank were used to train and evaluate a regularized logistic regression model for mortality and linear regression models for hospital length of stay (HLOS) and intensive care unit length of stay (ILOS) using ICD-9 diagnostic and procedural codes. Model training was performed on data from 2008 (n = 124,625) and evaluation on data from 2009 (n = 120,079). The discrimination and calibration of each model based on ICD-9 codes were compared with those of TRISS. RESULTS The mortality model using ICD-9 codes was comparable with that of TRISS in terms of the area under the receiver operating characteristic curve (0.922 versus 0.921, P = not significant.) and achieved better results in terms of both integrated discrimination improvement (0.106, P < 0.001) and Hosmer-Lemeshow chi-squared value (294.15 versus 2043.20). The HLOS and ILOS models using ICD-9 codes also demonstrated improvements in both R(2) (0.64 versus 0.30 for HLOS, 0.68 versus 0.34 for ILOS) and root mean-squared error (7.06 versus 8.62 for HLOS, 4.15 versus 9.54 for ILOS). CONCLUSIONS Use of ICD-9 codes for stratification provides a more accurate and more broadly applicable approach to quality and outcomes assessment in trauma patients than the labor-intensive gold standard of TRISS.