Hospital

Introduction

Since this is my first datathon I decided to attempt a challenge that I thought was possible for the minimum level of knowledge I had.

What it does

I have built a small neural network that effectively can predict if a hospital patient will live based on 43 different categories.

How we built it

Over the past day, I have learned how to build a neural network and use various data points to train the network. During my journey, I would attempt to change one thing at a time and see if that would increase the percentage of accuracy.

Challenges we ran into

When I started I was trying to logically think through what columns should be most likely linked to the deaths of patients. The problem with that approach is that I was limiting the amount of data that the neural network had access to by about a quarter and the accuracy was capped at around .7. Next, I systematically changed the columns with terms in them into numbers. By doing this, I increased the amount of data from around 11 columns to 43 columns. At this point, the accuracy became around .9. although it varied greatly and I noticed that I was overfitting since my train data was having an ever increasing accuracy (from .95 to .99) while my validation and test were hanging around .9. To fix this I added and tested many different regularizers that would keep my training data to having a similar accuracy to my validation and test accuracy. This would consistently give me an accuracy of more than .9. These were only the effective test I had though, I tried changing the number of neurons in the hidden layers and honed in on the fact that the number of neurons should be roughly equal to the number of inputs. I also tried different batch sizes and number of epochs and realized that a batch size of 16 was the goldy locks of batch sizes and the number of epochs should be around 20 as 10 was still becoming more accurate and after 20 it would just keep bouncing around .9. There were other things that I tested though these were the ones with the greatest effect on the accuracy of my neural network.

Accomplishments that we're proud of

Although it took quite a while I am quite proud of all I learned in the time that I have put into this project.

What's next for Hospital

The next things that I will be attempting is to make a random tree and possibly combine the random tree with the model I created to get a higher accuracy.

Built With

• numpy
• pandas
• python
• scikit-learn
• tensorflow