Subramani Mani

A population-based analysis of specific behavior problems associated with childhood seizures

Abstract A population-based analysis was conducted, with the 1988 National Health Interview Survey (NHIS) to estimate the risk for behavior problems associated with children with seizures and cardiac conditions as compared with that of children with no known health conditions. The Behavior Problem Index (BPI) was used to compare the frequency of problems and the differences in scores between groups, using parent responses to a 28-item behavior checklist. In addition, the odds of having a behavior problem in the three groups of children with health conditions were compared with those of children with no known health conditions. Our analysis indicates that children with seizures were 4.7 times more likely to have a behavior problem and children with cardiac problems were 3 times more likely to have behavior problems as compared with controls. The specific behaviors and traits that were most problematic for children with seizures were being hyperactive (cannot concentrate, easily confused, impulsive, obsessive, and restless) and dependent (clings to adults, cries a lot, demands a lot of attention, too dependent on others).

Medical decision support using machine learning for early detection of late-onset neonatal sepsis

OBJECTIVE The objective was to develop non-invasive predictive models for late-onset neonatal sepsis from off-the-shelf medical data and electronic medical records (EMR). DESIGN The data used in this study are from 299 infants admitted to the neonatal intensive care unit in the Monroe Carell Jr. Childrens Hospital at Vanderbilt and evaluated for late-onset sepsis. Gold standard diagnostic labels (sepsis negative, culture positive sepsis, culture negative/clinical sepsis) were assigned based on all the laboratory, clinical and microbiology data available in EMR. Only data that were available up to 12 h after phlebotomy for blood culture testing were used to build predictive models using machine learning (ML) algorithms. MEASUREMENT We compared sensitivity, specificity, positive predictive value and negative predictive value of sepsis treatment of physicians with the predictions of models generated by ML algorithms. RESULTS The treatment sensitivity of all the nine ML algorithms and specificity of eight out of the nine ML algorithms tested exceeded that of the physician when culture-negative sepsis was included. When culture-negative sepsis was excluded both sensitivity and specificity exceeded that of the physician for all the ML algorithms. The top three predictive variables were the hematocrit or packed cell volume, chorioamnionitis and respiratory rate. CONCLUSIONS Predictive models developed from off-the-shelf and EMR data using ML algorithms exceeded the treatment sensitivity and treatment specificity of clinicians. A prospective study is warranted to assess the clinical utility of the ML algorithms in improving the accuracy of antibiotic use in the management of neonatal sepsis.

Applying active learning to assertion classification of concepts in clinical text

Supervised machine learning methods for clinical natural language processing (NLP) research require a large number of annotated samples, which are very expensive to build because of the involvement of physicians. Active learning, an approach that actively samples from a large pool, provides an alternative solution. Its major goal in classification is to reduce the annotation effort while maintaining the quality of the predictive model. However, few studies have investigated its uses in clinical NLP. This paper reports an application of active learning to a clinical text classification task: to determine the assertion status of clinical concepts. The annotated corpus for the assertion classification task in the 2010 i2b2/VA Clinical NLP Challenge was used in this study. We implemented several existing and newly developed active learning algorithms and assessed their uses. The outcome is reported in the global ALC score, based on the Area under the average Learning Curve of the AUC (Area Under the Curve) score. Results showed that when the same number of annotated samples was used, active learning strategies could generate better classification models (best ALC-0.7715) than the passive learning method (random sampling) (ALC-0.7411). Moreover, to achieve the same classification performance, active learning strategies required fewer samples than the random sampling method. For example, to achieve an AUC of 0.79, the random sampling method used 32 samples, while our best active learning algorithm required only 12 samples, a reduction of 62.5% in manual annotation effort.

Two-Stage Machine Learning model for guideline development

We present a Two-Stage Machine Learning (ML) model as a data mining method to develop practice guidelines and apply it to the problem of dementia staging. Dementia staging in clinical settings is at present complex and highly subjective because of the ambiguities and the complicated nature of existing guidelines. Our model abstracts the two-stage process used by physicians to arrive at the global Clinical Dementia Rating Scale (CDRS) score. The model incorporates learning intermediate concepts (CDRS category scores) in the first stage that then become the feature space for the second stage (global CDRS score). The sample consisted of 678 patients evaluated in the Alzheimers Disease Research Center at the University of California, Irvine. The demographic variables, functional and cognitive test results used by physicians for the task of dementia severity staging were used as input to the machine learning algorithms. Decision tree learners and rule inducers (C4.5, Cart, C4.5 rules) were selected for our study as they give expressive models, and Naive Bayes was used as a baseline algorithm for comparison purposes. We first learned the six CDRS category scores (memory, orientation, judgement and problem solving, personal care, home and hobbies, and community affairs). These learned CDRS category scores were then used to learn the global CDRS scores. The Two-Stage ML model classified as well as or better than the published inter-rater agreements for both the category and global CDRS scoring by dementia experts. Furthermore, for the most critical distinction, normal versus very mildly impaired, the Two-Stage ML model was 28.1 and 6.6% more accurate than published performances by domain experts. Our study of the CDRS examined one of the largest, most diverse samples in the literature, suggesting that our findings are robust. The Two-Stage ML model also identified a CDRS category, Judgment and Problem Solving, which has low classification accuracy similar to published reports. Since this CDRS category appears to be mainly responsible for misclassification of the global CDRS score when it occurs, further attribute and algorithm research on the Judgment and Problem Solving CDRS score could improve its accuracy as well as that of the global CDRS score.

Pharos: Collating protein information to shed light on the druggable genome

The ‘druggable genome’ encompasses several protein families, but only a subset of targets within them have attracted significant research attention and thus have information about them publicly available. The Illuminating the Druggable Genome (IDG) program was initiated in 2014, has the goal of developing experimental techniques and a Knowledge Management Center (KMC) that would collect and organize information about protein targets from four families, representing the most common druggable targets with an emphasis on understudied proteins. Here, we describe two resources developed by the KMC: the Target Central Resource Database (TCRD) which collates many heterogeneous gene/protein datasets and Pharos (https://pharos.nih.gov), a multimodal web interface that presents the data from TCRD. We briefly describe the types and sources of data considered by the KMC and then highlight features of the Pharos interface designed to enable intuitive access to the IDG knowledgebase. The aim of Pharos is to encourage ‘serendipitous browsing’, whereby related, relevant information is made easily discoverable. We conclude by describing two use cases that highlight the utility of Pharos and TCRD.

Detecting Very Early Stages of Dementia from Normal Aging with Machine Learning Methods

We used Machine Learning (ML) methods to learn the best decision rules to distinguish normal brain aging from the earliest stages of dementia using subsamples of 198 normal and 244 cognitively impaired or very mildly demented (Clinical Dementia Rating Scale=0.5) persons. Subjects were represented by their age, education and gender, plus their responses on the Functional Activities Questionnaire (FAQ), the Mini-Mental Status Exam (MMSE), and the Ishihara Color Plate (ICP) tasks. The ML algorithms applied to these data contained within the electronic patient records of a medical relational database, learned rule sets that were as good as or better than any rules derived from either the literature or from domain specific knowledge provided by expert clinicians. All ML algorithms for all runs found that a single question from the FAQ, the forgetting rule, (“Do you require assistance remembering appointments, family occasions, holidays, or taking medications?”) was the only attribute included in all rule sets. CARTs tree simplification procedure always found that just the forgetting rule gave the best pruned decision tree rule set with classification accuracy (93% sensitivity and 80% specificity) as high as or better than any other decision tree rule-set. Comparison with published classification accuracies for the FAQ and MMSE revealed that including some of the additional attributes in these tests actually worsen classification accuracy. Stepwise logistic regression using the FAQ attributes to classify dementia status confirmed that the forgetting rule gave a much larger odds ratio than any other attribute and was the only attribute included in all of the stepwise logistic regressions performed on 33 random samples of the data. Stepwise logistic regression using the MMSE attributes identified two attributes which occurred in all 33 runs and had by far the highest odds ratio. In summary, ML methods have discovered that the simplest and most sensitive screening test for the earliest clinical stages of dementia consists of a single question, the forgetting rule.

MENTOR: A bayesian model for prediction of mental retardation in newborns

Mental retardation (MR) is a diagnosis that is made with extreme caution because of the many uncertainties in its etiology and prognosis. In fact, most physicians will delay the diagnosis for months or years so that substantial evidence is available to rule the diagnosis in or out. MENTOR is a Bayesian Model for the prediction of MR in newborns that provides probabilities for the full range of cognitive outcomes, ranging from MR to superior intelligence. Using the model to confirm clinical judgment could help physicians decide when to proceed with diagnostic tests. The physician and family could discuss the probabilities for MR, borderline, normal, and superior intelligence, given the childs status in infancy and base their decision about additional testing, in part, on this information.

Knowledge Discovery from a Breast Cancer Database

We report on the use of various Machine Learning algorithms on an electronic database of breast cancer patients. The task was to predict breast cancer recurrence using a short subset of clinical attributes such as tumor presence, tumor size, invasive nature of tumor, number of lymph nodes involved, severity of lymphedema and stage of tumor. The predictive accuracy over fifty runs employing test sets not used to build the model were 63.4%(Cart), 63.9%(C45), 62.5%(C45rules), 66.4%(FOCL) and 68.3%(Naive Bayes). An extension of the model using additional features and larger datasets is contemplated.

Machine learning for predicting the response of breast cancer to neoadjuvant chemotherapy

OBJECTIVE To employ machine learning methods to predict the eventual therapeutic response of breast cancer patients after a single cycle of neoadjuvant chemotherapy (NAC). MATERIALS AND METHODS Quantitative dynamic contrast-enhanced MRI and diffusion-weighted MRI data were acquired on 28 patients before and after one cycle of NAC. A total of 118 semiquantitative and quantitative parameters were derived from these data and combined with 11 clinical variables. We used Bayesian logistic regression in combination with feature selection using a machine learning framework for predictive model building. RESULTS The best predictive models using feature selection obtained an area under the curve of 0.86 and an accuracy of 0.86, with a sensitivity of 0.88 and a specificity of 0.82. DISCUSSION With the numerous options for NAC available, development of a method to predict response early in the course of therapy is needed. Unfortunately, by the time most patients are found not to be responding, their disease may no longer be surgically resectable, and this situation could be avoided by the development of techniques to assess response earlier in the treatment regimen. The method outlined here is one possible solution to this important clinical problem. CONCLUSIONS Predictive modeling approaches based on machine learning using readily available clinical and quantitative MRI data show promise in distinguishing breast cancer responders from non-responders after the first cycle of NAC.

Study of active learning in the challenge

In the active learning challenge, we aim to improve the area under the learning curve (ALC), the global score in the challenge, by optimizing the classification methods and feature selection methods, and most importantly by refining the querying algorithm to select the most informative instances in the early iteration of active learning. For six different datasets in the development phase, we applied general and specific methods to resolve unbalanced class, sparse data, and missing value problems. We designed a voting system based on multi models to combine good prediction with robust performance in different types of datasets. For querying methods, we modified the approach of information density, firstly, to avoid the exhaustive comparison for all samples, and secondly to find more representative samples. We also propose two modified versions of uncertainty sampling based methods: uncertainty sampling with bias, which takes into account the high imbalance of data, and uncertainty sampling with prediction, which predicts the most uncertain samples based on the change of uncertain values during the active learning process. We present our preliminary results of the development datasets in the active learning challenge and discuss their significance in this paper.