Outline of the Workshop
1. Setup and Python/Jupyter Notebook Basics
2. 10 minutes to Pandas/DataFrame (actually about 30 mins)

Amazon Web Services (AWS) provides cloud computing resources to companies around the world. These services include data storage, computing, and analytics. In November 2017, AWS released SageMaker. SageMaker is an end-to-end service which allows developers and data scientists to explore data, train models, and deploy models in production all within an easy to use platform. SageMaker comes with built-in common algorithms but also allows data scientists to bring their own algorithms in many languages including R and Python. During this presentation we will walk through a use case of SageMaker in the digital analytics industry. The use case will include all steps of the modeling process including data preparation, training, and deployment.

As healthcare continues to transition to value-based care, understanding how to effectively leverage data for population health is essential. Sanford Health recognizes this vital step and is leveraging models of innovation and collaboration to ensure success in a shifting healthcare climate. This talk will address current efforts including the Sanford Data Collaborative as well as introduce advanced analytics teamwork occurring across our footprint.

In presenting clinical data, Table 1 is used to present demographic, clinical, and pre-treatment data for trials. This includes continuous variables (age, BP, BMI - presented as means&stddev or median-IQR) and discrete variables (gender, race, cancer stage - presented as proportions-counts). While the actual statistical/data analytic techniques involved are trivial, constructing such tables takes a lot of time as results must be assembled into standard tables. Automating this is a huge time saver. SAS and R methods for building this table in a consistent manner are presented.

This presentation will outline the development and use of a probabilistic measure for outlier detection, referred to as atypicality. Given a set of objects, we can create a corresponding set of similarity scores between them. Assuming the set of scores has a normal distribution, we can estimate the score distribution’s parameters. We compute atypicality by comparing the likelihood of an object given these estimated parameters to the likelihood of bootstrapped samples. The atypicality measure is then used as a p-value in a hypothesis test, where the null hypothesis states that the object in question is similar to the remaining objects; the alternative hypothesis is that the object is an outlier. This can be used in a variety of applications, especially where we have multiple objects in multi-dimensional space.

Do you want to collect more dollars while spending less? Net lift analysis can help target customers with the highest propensity to be positively influenced by a collections treatment while balancing the costs of collecting. A collections effectiveness case study was conducted that outlines the basics of net lift analysis. Strategies using net lift analysis resulted in an increase in dollars collected and a decrease in collections costs. Key takeaways are the importance testing, using the correct metrics, and evolving as you learn.

estimate the quantile elasticity of intertemporal substitution (QEIS) of consumption using instrumental variables quantile regression. The elasticity of intertemporal substitution represents the willingness of a consumer to substitute future consumption for present consumption. In this paper, agents have a quantile utility preference instead of standard expected utility. This allows for the capture of heterogeneity along the conditional distribution of agents. The QEIS considers structural breaks in the data and is estimated for each regime using linearized Epstein-Zin preferences and by the use of fixed effects, instrumental variables, and quantile regression. The estimator is a feasible estimator based on smoothed sample moments. In order to estimate the model, the Nielsen Consumer Panel dataset is used. This dataset is built from transactional data that follows households in the United States and their grocery purchases from 2004 through 2014. Because of the transactional nature of the dataset, there is a low source of measurement error in consumption, and aggregation bias can be minimized. To estimate the model, consumption is aggregated weekly, and consumption growth is measured over a four-week time period in order to match four-week Treasury bills. Results give evidence of heterogeneity of the QEIS along the quantiles of the conditional distribution.

It is known that a protein’s biological function is in some way related to its physical structure. Many researchers have studied this relationship both for the entire backbone structures of proteins as well as their binding sites, which are where binding activity occurs. However, despite this research, it remains an open challenge to predict a protein’s function from its structure. There are many useful applications from protein function predictions, such as effective drug discovery with fewer side effects, development of structure-based drug designs, disease diagnosis, and many more. This presentation will discuss how this ligand-binding protein prediction problem is approached by taking a higher level object-oriented approach, which is named as Covariances of Distances to Principal Axis (CDPA) that summarizes the description of the binding site so that it reduces the amount of information lost compared to most of the other approaches. Thereby, a model-based method is considered, where the nonparametric model is implemented by using the features of the binding sites for a given ligand group for understanding and classification purposes. Then the results obtained using the model-based approach are compared to the alignmentbased method used by Ellingson and Zhang (2012) and Hoffmann et al. (2010).

Patient portals are positioned as a central component of patient engagement through the potential to change the physician-patient relationship and enable chronic disease self-management. In this article, we extend the existing literature by discovering design gaps for patient portals from a systematic analysis of negative users’ feedback from the actual use of patient portals. Specifically, we adopt topic modeling approach, LDA algorithm, to discover design gaps from online low rating user reviews of a common mobile patient portal, EPIC’s mychart. To validate the extracted gaps, we compared the results of LDA analysis with that of human analysis. Overall, the results revealed opportunities to improve collaboration and to enhance the design of portals intended for patient-centered care.

Coupled systems of nonlinear reaction-diffusion PDE’s have been model pattern formation since the 1950’s. In particular, Turing proved that minor perturbations to initial conditions can result in exotic pattern formations. More recently, systems of PDE have been used to model plant/groundwater interactions. In this talk, we will discuss a recent work on mathematical models used to model the growth of vegetation on or around termite mounds in the African Savannah.

Rapid innovation in biotechnology, especially DNA sequencing, holds great promise for revolutionizing medicine. The main bottleneck is how to analyze and interpret the massive amount of data effectively. Many stand-alone software packages exists, mostly as R packages. We developed iDEP(Integrated Differential Expression and Pathway analysis), a large Shiny application that integrates hundreds of R packages, and large gene annotation database. Available at (http://ge-lab.org/idep/), iDEP streamlines complex bioinformatics pipelines as a friendly web interface, that can turn data into biological insights within minutes, instead of months.

It should come as no surprise that many data science and analytics projects fail. There are a whole number of reasons, and this talk will cover some of them. We will walk through the journey of planning a pizza party, studying for a test, and a few other fun stories. All of it will relate back to challenges with data science in the real world.

A career in marketing data science will entail answering questions about how best to allocate spend within a marketing portfolio. As marketing budgets continue to shift towards online channels and away from print, the availability and quantity of campaign data make the data scientist’s job more challenging than ever. Within financial services (and credit card issuers), a strategy that optimizes the mix of paid direct mail, paid online, organic search and on-site optimization has historically led to more efficient marketing costs. That strategy can been derived through attribution modeling, a close cousin to marketing mix models. In this presentation, Nirav will present examples of the conversion path of customers over paid, earned and owned marketing channels. He’ll introduce the data engineering and analytics tools used to give credit to each channel for conversions. And, he’ll introduce a few areas where science and tools are nascent.

I train a multiple-hidden-layer deep neural network that predicts self-perception of heart health (including Cardiovascular Disease) with a large data set of 1729 features and about 30,000 samples. The dataset is based on the CDC Demographics, Dietary, Examination, Laboratory, and Questionnaire datasets collected from 1999 to 2016. Substantial data cleaning and pre-processing are done with Python pandas library. The objective of this research is three-fold:

• Better understanding of how well DNN would improve the accuracy of prediction on perception of cardiovascular disease;

• Framework of developing more sophisticated DNN models to predict medical outcomes;

• Foundation for learning multi-dimensional/distributed representation of healthcare concepts that are both interpretable and scalable.

The presentation focuses on the technical (training a deep neural network with latest developments in the field of deep learning) aspects of the research.

Patients not showing up to their appointments is a detriment to both the patient and the health care system. As health care systems transition from fee-for-service programs to value-based program, clinical visits (especially primary care) will become the gateway to improving overall patient health outcomes. In order to ensure patients are receiving the appropriate treatment and maintaining a healthy lifestyle they must be completing their scheduled visits. The main goal of the model is to predict patient no-show probabilities with the intent of taking the model one step further by linking it to actionable data points and decisions. The model employs the use of a logistic regression and Bayesian update approach. The regression is devised of patient demographical, behavioral and diagnosis characteristics, as well as visit logistics. The logistic regression creates a priori probability based on requisite factors. Then due to the highly behavioral impetus of missing appointments, a Bayesian update is applied to the priori probability to obtain a final, posterior probability. The Bayesian application to this model significantly contributes to the patient’s probability and details the importance behind patient-level interventions. The output of the model has a high level of accuracy that allows clinics not only to see which patients have a high risk of not showing up, but also the factors that physicians may be able to remedy down the road. The model was built using a standard 10-fold cross-validation. The test set was then ran through the model and used to determine the weighting for the Bayesian update. Lastly the data was validated using the remaining 10%, which resulted in an AUC of .927. Combining the accuracy of this model with the prescriptive ability of the factors, can allow for a significant reduction of no-shows, not only by enhancing appointment logistics (calls, overbooking, etc.) but also by improving patients’ lifestyle.

The Polaris data science team was founded with a simple objective: be predictive. Driving business value with initial quick wins was relatively easy, but scaling our ability to drive change was limited by the small size of the team. Furthermore, many of the business units needed fundamental reporting automation more than a neural network. We will discuss how we pivoted our strategy and are taking a more holistic self-service enablement approach alongside predictive algorithm development.

Background In the era of big data, a large number of clinical narratives exist in electronic health records. Automatic extraction of key variables from clinical narratives has facilitated many aspects of healthcare and biomedical research. Conventional approaches are based on rule-based natural language processing (NLP) techniques that rely on expert knowledge and exhaustive human efforts of designing rules. Recently machine learning has seen a big performance gain compared to conventional NLP approaches. Despite the impressive improvements achieved by machine learning models, large manual labeled training data are the crucial building blocks of conventional machine learning methods and key enablers of recent deep learning methods. However, large training data are not always readily available and usually expensive to obtain from human annotators. This problem becomes more significant for use cases in clinical domain due to the Health Insurance Portability and Accountability Act (HIPAA) where methods, such as crowdsourcing, are not applicable, and requirements of annotators being medical experts. Method In this paper, we propose a distant supervision paradigm for clinical information extraction. In this paradigm, rule-based NLP algorithms are used to generate large training data with labels automatically. Machine learning models are subsequently trained on these distant labels with word embedding features. Results We study the effectiveness of the proposed framework on two clinical information extraction tasks i2b2 smoking status extraction shared task and a fracture extraction task at our institution. We tested three prevalent machine learning models, namely, Convolutional Neural Networks, Support Vector Machine, and Random Forrest. Conclusion The experimental results show that the proposed distant supervision paradigm is effective for the machine learning models to learn rules towards gold standard from distant labels. Moreover, the machine learning models trained on the distant labels generated by a rule-based NLP algorithm could perform better than the NLP algorithm given sufficient data. Additionally, we showed that CNN was more sensitive to the data size than the conventional machine learning models and that all the tested machine learning methods were viable options for the distant supervision paradigm.

Variable selection is an important means to construct a model that predicts a target/responsible variable with a set of predictable variables. The predictable variables could include quantitative, binary, and/or categorical variables; however, commonly used variable selection methods such as forward selection, backward selection, and stepwise selection are more focused on quantitative variables. It will be a helpful addition to multiple linear regression if categorical variables can be integrated with the commonly used variable selection methods. We proposed a generalized variable selection method that can be used to select both categorical and quantitative variables simultaneously. The detailed results will be presented at the symposium.

This talk will share the details of an intriguing and appealing in-class project for students in an introductory Data Analytics class for advanced mathematics majors. The project originated with a joke from a popular comedian about whether “La Quinta” is Spanish for “Next to Denny’s” and developed into an investigation of that quip. In this project, students learn to acquire the appropriate geospatial data, learn some new skills in Excel, RStudio or other data analysis software, experience quite a bit of problem-solving, and work hard to communicate their results.

A few years ago, I was contacted by Rodney Oppegard, a church historian. He had spent many years collecting information on ecclesiastical furnishings and artwork found in the Lutheran churches of rural North Dakota, and while his data set was extensive it was by no means complete. Some artwork was unsigned or the signature obscured, other pieces had been transferred to different churches, and in some cases the church itself had been destroyed by fire years before, leaving only incomplete records and fading memories as clues to the original church’s configuration. Mr. Oppegard wanted to know whether there was a mathematical way to use existing data to “fill in the holes” of his data set. In this talk, I will outline how geospatial and rudimentary archival data were used to construct and evaluate models for determining which of several popular artists was responsible for a particular church’s altar painting.

In the new era of development, social media is not only getting popularity but also useful to reveal hidden information. Most of the people use social media to connect friends and family, to express their emotions, to give feedback, and to raise concerns as quickly as possible. We can reveal important information from people responses in social media. OSCAR nominations were announced for the year 2017 and all the nominees are very active on Twitter. All the tweets regarding their nominations are publicly available. Here, in this paper, we analyzed the public tweets from Twitter and predict who will be the OSCAR winner in 2017. More specifically, we analyzed all the tweets of the OSCAR nominees in the category of “Actor in leading role” since the day of OSCAR nomination announcement. After analyzing the data, we predicted who will be the winner based on the twitter sentiment analysis using R programming.

In this talk, a set of procedures for building R package will be discussed and some recently built R packages will be introduced.

Social media generates huge amounts of data every day and most of the data is unstructured. This is a untapped resource that may provide significant benefits to companies able to exploit this data. SAS Visual Analytics is a big data tool that facilitates the visualization of large data sets. In this talk we demonstrate how insight can be derived from the analysis of Donald Trump’s tweets. First, a word cloud is built and then a sentiment analysis is done on all of his tweets. Tweets are grouped into topics and the sentiment surrounding each topic is analyzed. This leads to the discovery of novel and interesting insights.