Qing Zeng-Treitler

Abstracts

3 Shedding Light on the Black Box: Explaining Deep Neural Network Prediction of Clinical Outcome

Authors: Qing Zeng-Treitler, Yan Cheng, Yijun Shao, Charlene R. Weir, Rashmee U. Shah, Bruce E. Bray

Abstract:

Deep neural network (DNN) models are being explored in the clinical domain, following the recent success in other domains such as image recognition. For clinical adoption, outcome prediction models require explanation, but due to the multiple non-linear inner transformations, DNN models are viewed by many as a black box. In this study, we developed a deep neural network model for predicting 1-year mortality of patients who underwent major cardio vascular procedures (MCVPs), using temporal image representation of past medical history as input. The dataset was obtained from the electronic medical data warehouse administered by Veteran Affairs Information and Computing Infrastructure (VINCI). We identified 21,355 veterans who had their first MCVP in 2014. Features for prediction included demographics, diagnoses, procedures, medication orders, hospitalizations, and frailty measures extracted from clinical notes. Temporal variables were created based on the patient history data in the 2-year window prior to the index MCVP. A temporal image was created based on these variables for each individual patient. To generate the explanation for the DNN model, we defined a new concept called impact score, based on the presence/value of clinical conditions’ impact on the predicted outcome. Like (log) odds ratio reported by the logistic regression (LR) model, impact scores are continuous variables intended to shed light on the black box model. For comparison, a logistic regression model was fitted on the same dataset. In our cohort, about 6.8% of patients died within one year. The prediction of the DNN model achieved an area under the curve (AUC) of 78.5% while the LR model achieved an AUC of 74.6%. A strong but not perfect correlation was found between the aggregated impact scores and the log odds ratios (Spearman’s rho = 0.74), which helped validate our explanation.

Keywords: prediction, logistic regression model, frailty, deep neural network, temporal data

Procedia PDF Downloads 23
2 Experiments on Weakly-Supervised Learning on Imperfect Data

Authors: Qing Zeng-Treitler, Yan Cheng, Yijun Shao, James Rudolph, Charlene R. Weir, Beth Sahlmann

Abstract:

Supervised predictive models require labeled data for training purposes. Complete and accurate labeled data, i.e., a ‘gold standard’, is not always available, and imperfectly labeled data may need to serve as an alternative. An important question is if the accuracy of the labeled data creates a performance ceiling for the trained model. In this study, we trained several models to recognize the presence of delirium in clinical documents using data with annotations that are not completely accurate (i.e., weakly-supervised learning). In the external evaluation, the support vector machine model with a linear kernel performed best, achieving an area under the curve of 89.3% and accuracy of 88%, surpassing the 80% accuracy of the training sample. We then generated a set of simulated data and carried out a series of experiments which demonstrated that models trained on imperfect data can (but do not always) outperform the accuracy of the training data, e.g., the area under the curve for some models is higher than 80% when trained on the data with an error rate of 40%. Our experiments also showed that the error resistance of linear modeling is associated with larger sample size, error type, and linearity of the data (all p-values < 0.001). In conclusion, this study sheds light on the usefulness of imperfect data in clinical research via weakly-supervised learning.

Keywords: Simulation, prediction, Delirium, support vector machine, weakly-supervised learning

Procedia PDF Downloads 28
1 Incorporating Information Gain in Regular Expressions Based Classifiers

Authors: Rosa L. Figueroa, Christopher A. Flores, Qing Zeng-Treitler

Abstract:

A regular expression consists of sequence characters which allow describing a text path. Usually, in clinical research, regular expressions are manually created by programmers together with domain experts. Lately, there have been several efforts to investigate how to generate them automatically. This article presents a text classification algorithm based on regexes. The algorithm named REX was designed, and then, implemented as a simplified method to create regexes to classify Spanish text automatically. In order to classify ambiguous cases, such as, when multiple labels are assigned to a testing example, REX includes an information gain method Two sets of data were used to evaluate the algorithm’s effectiveness in clinical text classification tasks. The results indicate that the regular expression based classifier proposed in this work performs statically better regarding accuracy and F-measure than Support Vector Machine and Naïve Bayes for both datasets.

Keywords: text classification, Smith-Waterman algorithm, information gain, regular expressions

Procedia PDF Downloads 152