Artificial Intelligence

in Intensive
Care Units

Deep Learning-Based Clinical Decisioning Tools

Challenge Description

Clinicians need reliable clinical decisioning tools to avert unexpected events that might harm their patients in intensive care units (ICU). Early prediction of unusual length of stay(LOS) and in-hospital mortality(IHM) for patients in Intensive Care Units (ICU) can help caregivers identify patients at high risk and effectively manage their care to prevent adverse events. Digitized healthcare data has been growing at a rapid pace making its productive utilization a challenging endeavor.

In this study, we proved that deep learning-based predictive models built using all available patient data can predict IHM and LOS with higher performance than models built using a subset of patient data. In order to build models that would learn from the entirety of Electronic Health Records (EHR) data, Natural Language Processing (NLP) techniques and deep learning neural networks were applied on MIMIC-III v1.4 (Medical Information Mart for Intensive Care III) dataset.

Approach

01

We utilized Electronic Health Records (EHRs) like chart data, input and output events, laboratory test results, caregiver notes, medications, procedures, imaging reports to build Deep Learning Models (DLMs).

02

Traditional methodologies for building Machine Learning models utilize curated set of variables, leaving majority of data unutilized. In contrast, our approach entailed usage of all digitally captured patient information from EHRs and other sources.

03

Natural Language Processing (NLP) techniques were extensively used to synthesize and combine structured and unstructured datasets.

04

A custom-built deep learning network with input layer, embedding layer, pooling layer, dense layer and output layer was used in the Architecture.

05

Used scalable cloud infrastructure of Google Cloud Platform (GCP) and its tools such as Cloud Storage, Big Query, TensorFlow, cloud functions and other Artificial Intelligence Platform tools to develop DLMs.

Results

In Hospital Mortality

AUC – ROC

PR-AUC

Balanced Accuracy

Length of Stay

AUC – ROC

PR-AUC

Balanced Accuracy

Conclusion

With this approach, we observed promising model performance from only 42,818 admissions. With a higher volume of patients, more variety of clinical data, and deep learning, one may achieve better model performance. In this study, we recognized the potential to develop the next generation of clinical practice tools that can augment the efficiency and effectiveness of caregivers in a critical care setting. However, it is only possible when the data is digitized and ready to be used for model development.

Publications

WANT TO LEARN MORE?