Modeling on a public-use Medicare dataset to attempt to predict several chronic health outcomes: diabetes, depression, and high-cost.
Medicare is the United States’ national social health insurance program for older adults. Medicare serves approximately 50 million Americans, the majority of which are 65 and older (some younger individuals with disabilities are also served). Machine learning can be applied in healthcare to identify patients at risk for specific health outcomes. We analyzed a publicuse Medicare dataset to attempt predictive modeling for several outcomes: diabetes, depression, and highcost. We selected diabetes due to its widespread prevalence in the United States (much of it undiagnosed) , and the significant economic toll it takes 1 on our healthcare system2. We chose depression because its cooccurrence with physical health conditions has been well documented3, and we wanted to see if we could predict its occurrence based on physical and demographic features in our dataset. We sought to predict highcost individuals because we believe a predictive algorithm for the costliest patients would be useful to insurance companies wishing to curb expenditures via preventive programs. We also initially attempted predictive models for death and endstage renal disease, although we abandoned these efforts for reasons we will address in a following section. Our hypothesis was that we could use machine learning models to effectively predict diabetes, depression, and highcost patient outcomes.
We loaded the dataset into a PostgreSQL database, and served it to Jupyter notebooks via ODBC connection. Initial preprocessing was performed on the fly while loading in the data, by way of SQL query. A death indicator was constructed by checking for patients with a listed death date in 2009 or 2010, so that 2008 data could be used predictively. We derived an age field from the listed date of birth, and we aggregated cost fields into subtotals for inpatient, outpatient, carrier claims, and total costs.
We ran four predictive models on our data. First, we constructed a naive Bayes classifier mixed model to accommodate both discrete and continuous inputs, adapting code provided in class; we used a training set to construct a probability distribution to apply against a test set. Next, we ran our data through a linear support vector machine (SVM) using sklearn, using a validation set to identify an optimal C value. We then applied a neural network classifier against our training and test set, using code furnished by the course instructor. We additionally ran a random forest model, using sklearn. Finally, we constructed a simple heuristic model from scratch. To simplify evaluation of our various models, we created a module called evaluate.py to output a set of standarized set of evaluation results. Specifically, this function calculates outputs the following: true positives, false positives, true negatives, false negatives, accuracy, sensitivity/recall, specificity, precision, and F1 score. Additionally, for reference, it also outputs the true rate of the outcome of interest in the sample. Values we paid close attention to are sensitivity out of all patients with diabetes, how many did the classifier classify correctly and precision out of the individuals whom the model predicted diabetes, how many were classified correctly.
Before applying the features into the models, we created dummy data for the demographic data and rescaled the medical expense data through whitening for both the linear support vector machine and the random forest classifier. In addition, we sampled 10% of the entire dataset to enable the SVM workable in our personal computers where computing resource is not enough to deal with all data. Before running the Random Forest Classifier, we also converted the continuous data coverage, inpatient, outpatient and carrier costs into binary variables by splitting the data at the median values. In testing our models, we employed a 67%/33% training/test split.
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent