Hospital Readmission Rates

# Hospital Readmission Rates
## Solving Patient Needs with Big Data and Machine Learning
### Joshua Goldberg <br><span style="font-size: 50%;">Big Data Platforms<br>Master of Science in Data Science<br>Winter 2019</span>

---

# Problem Statement

Hospital costs are rising partially because of high readmission rates within 30 days of patient release. Readmission rates have long been a trusted measure of effective and responsible care and have become a primary assessment driver in the healthcare industry.
---

# The Goal

The goal is to research and design a Big data solution that will meet the patient analytics needs of a large health care provider with 500,000 customers (patients) around the world. The big data analytics platform would identify at-risk patients based on past history, chart information, and patient trends. The provider should be able to use this data to identify at-risk patients and provide the necessary care to reduce readmission rates.
---

# Data

Data would be refreshed at numerous intervals, including at patient admission, treatment, and exit.

<table>
 <thead>
  <tr>
   <th style="text-align:left;"> Personal Information </th>
   <th style="text-align:left;"> Medical Information </th>
   <th style="text-align:left;"> Hospital Records </th>
   <th style="text-align:left;"> Financial Information </th>
  </tr>
 </thead>
<tbody>
  <tr>
   <td style="text-align:left;"> Date of birth </td>
   <td style="text-align:left;"> Blood group </td>
   <td style="text-align:left;"> Date of admission </td>
   <td style="text-align:left;"> Insurance provider </td>
  </tr>
  <tr>
   <td style="text-align:left;"> Contact information </td>
   <td style="text-align:left;"> Medical history </td>
   <td style="text-align:left;"> Cause of illness </td>
   <td style="text-align:left;"> Bill amount </td>
  </tr>
  <tr>
   <td style="text-align:left;"> Next in kin </td>
   <td style="text-align:left;"> Test reports/Medical images </td>
   <td style="text-align:left;"> Duration of stay </td>
   <td style="text-align:left;"> Insurance claims </td>
  </tr>
  <tr>
   <td style="text-align:left;"> Demographics </td>
   <td style="text-align:left;"> Hospital records </td>
   <td style="text-align:left;"> Doctor-in-charge </td>
   <td style="text-align:left;"> Credit related to medical history </td>
  </tr>
  <tr>
   <td style="text-align:left;"> Family </td>
   <td style="text-align:left;"> Admission history </td>
   <td style="text-align:left;"> Treatments/surgeries </td>
   <td style="text-align:left;"> Employment details </td>
  </tr>
</tbody>
</table>

---

# Analytical Solution

Several models will be deployed that provide probabilistic outcomes for patients around readmission, follow-up, and optimal patient care:

* A prediction score for patient readmission prior to admission
--

* A prediction score for readmission as patient receives care at the hospital
--

* A predictive model that determines a patient’s at-risk at the time of exit.
--

* Clustering model that groups patients based on the propensity to re-admit; this model would help understand characteristics better of these two groups and can be used when the other predictive models disagree as a subjective tie-breaker
--

* Once the at-risk patience are identified, a model will be deployed that provide follow-up recommendations to help improve readmission rates
--

Analytics would be performed at numerous points in time, including patient entry, daily/intermittently as medical records are updated, patient exit (throughout the extent of possible readmission classification
---

# Big Data Technology

Data Storage format and compression technique
* ORC stores collections of rows in one file and within the collection the row data is stored in a columnar format. This allows parallel processing of row collections across a cluster.  It uses specific encoders for different column data types to improve compression further.

Database and Query Execution Engine
* Hive for SQL like queries – as you can simply map HDFS files to Hive tables and query the data. Even the HBase tables can be mapped and Hive can be used to operate on that data.
* Hbase for real-time querying of data. It is used if the application requires random read or random write operations or both.

---

# Big Data Technology

Analytics and Data Science Platform
* Apache spark can be used for machine learning and analytics – powerful unified engine, machine language are supported, Apache Spark is one of the most actively developed open source platforms

Cloud vs. On premise decision
* Hosting it on the cloud may cost more if there is consistent long term usage
* It may cost less to actually build your own Hadoop cluster on premise

---

# Estimating Data Capacity Requirements

* Number of patients: `\(500,000\)`
* Average size of electronic medical records including images: `\(80MB\)`
* Assuming that the readmission rate: `\(20\%\)`
* Growth of data: `\(20\%\)`

Total starting size: `\(500,000 \times80MB=0.04PB\)`, or about `\(150TB\)`

---

# Cluster/Node Capacity

* `\(H=RCS\times(1+T)\times(1+G)\)`
* Assuming `\(T=10%\)`, `\(G=20%\)`, `\(R=3\)`, `\(C=1\)`, `\(S=0.04PB\)`
* `\(H=3\times1\times0.04\times(1.1)\times(1.2)=0.1584PB= 158.4 TB\)`
* Assuming each node has a capacity of 24 TB

`\(n=\frac{H}{d}=\frac{1584}{24}=6.6\)`; therefore 7 nodes for initial data

* R: Replication factor. Usually 3 in a production cluster.
* C: Compression ratio. When no compression is used, `\(C=1\)`.
* S: Initial size of data that needs to be moved to HDFS.
* G: Data growth factor 
* T: Temporary Space; Usually 10-25% of working space
* d: size of each data node

---

# Monthly Cost on Average

* Hardware cost (7x per node cost): `\(63,000\)` 
* Software cost (7x per node cost): `\(28,000\)`
* Environment, Power, Cooling, etc: `\(31,000\)`
* Full time employees (1x): `\(110,000\)`

Total cost: `\(232,000\)`<br>
Approximate monthly cost: `\(20,000\)`

---

<h1 style="text-align: left;"> Architecture Diagram</h1>

![](slide-content/big-data-architecture.svg)