Clinical Decision Making Using Machine Learning and ICU Data

Author Name: Saurabh P. Tembhurne, Dr. Shubhangi Neware
Author Email: tembhurnesp@rknec.edu

Abstract

Clinical decision making using machine learning algorithms, is a complicated task as it requires more accurate results in less time. Since the available amount of data is sparse and heterogeneous the design of such decision making systems requires relevant and the most suitable machine learning algorithm. In proposed research work we utilized the enormous capabilities of Recurrent Neural Network (RNN) which are proven to be very successful in predicting patients health related event/condition that may occur in future. The prime focus of proposed research work is to predict the disease which can occur with the help of selected features. We have considered and analysed two approaches RNN with LSTM and RNN with RELU for their effectiveness and practical adequacy particularly on the healthcare datasets. Proposed research work gives a conclusion that the performance of the above machine learning techniques will depend on the number of features selected and the way in which datasets are analysed with more precision on the health care dataset. Accuracy and speed of algorithm is considered as important performance parameter for this proposed work. RNN with RELU are producing promising results for making clinical decisions by predicting patient’s future condition using current features of patient’s dataset.

Keywords

Recurrent Neural Networks (RNN), Machine Learning, Feed Forward Neural Network, Rectified Linear Unit (RELU), Long Short Term Memory (LSTM) and Gated Recurrent Unit (GRU).

Introduction

The basis of decisions by doctors in clinical working environment or while treating patients are mostly based on the past record of patients health and medical knowledge they have. But now in the era of digitization vast amount of medical knowledge and information hide inside the medical records inside the hospital, genomic databases and scientific publications. As the recent trend is towards digitization the information in these records are rapidly growing day by day. Due to this human experts are getting deluged from this huge amount of information. Healthcare is in the middle of this huge amount of “Data” explosion. That’s why Machine Learning is gaining attention in the healthcare domain from past few years. Machine learning take advantage of this information contained in these vast data sources to predict the most likely future events that may occur. These predictions helps physician in their decision making process while treating to their patients, which may lead to the better results.

A class of algorithms that can model very complex relationships are Neural Networks, which have proven to be very successful in other areas of Machine Learning [3]. Particularly, there is notable similarity among the prediction of clinical events and the field of Language Modelling, where Deep Learning, a class of Neural Networks with multiple hidden layers, has also proven to be very successful. One could imagine that each word of a text represents an event. Therefore a text would be a stream of events and the task of Language Modelling would be to predict the next event in the stream [1]. However, the clinical area has a set of unique features that make it an almost different scenario: more than one events can occur at a time, there are multiple patients, each sequence has an associated set of fixed information like patient name, age, gender, blood group, etc. Because of this we need to develop new model especially for the medical dataset.

Conclusion

We developed and compared two algorithms based on RNNs that are efficient for clinical dataset, in order to predict various diseases which can occur during lifespan of patient. This will provide physicians an insight about patient’s next state earlier and will backing them for better decision making. We found that an RNN with RELU along with a Feedforward Neural Network gives the best result for this task. We also compared our model with RNN composed of LSTM and found that our proposed model outperform these recurrent model on the basis of accuracy and time requirement.

85 total views, 1 views today

Download File

About the author: admin