Outline

1. Introduction
2. Process overview
3. Notebook descriptions
4. Input data

1. Introduction

This repository contains code, predictions and logs for the manuscript that has been submitted for publication as:

Menger, V., Spruit, M., van Est, R., Nap, E., Psydata group, Scheepers, F. (2019). Inpatient violence risk assessment using routinely collected clinical notes in Electronic Health Records: a machine learning approach.

The analysis code consists of a collection of Jupyter notebooks that preprocess data, perform exploratory analysis and machine learning modeling, and evaluate outcomes using Python. Additionally, all output logs and predictions for both sites are available in the logs folder.

2. Process overview

A schematic depiction of the relations between data, notebooks and trained models can be seen below.

3. Notebook descriptions

The following notebooks are included in this repository.

Notebook	Description
train_paragraph2vec_model.ipynb	Processes the texts in the notes_full dataset, and then trains a paragraph2vec model.
integrate_dataset.ipynb	Processes and integrates the admissions, incidents and notes datasets.
exploratory_analysis.ipynb	Performs exploratory analysis on the integrated dataset.
train_classification_model.ipynb	Trains a classification model using the integrated dataset, and produces an internal prediction using cross validation.
evaluate_models.ipynb	Evaluates model performance, both the cross validated prediction, and a prediction using models that were externally trained on a similar dataset.

4. Input data

The scripts assume as input the four data tables with specified columns.

1. admissions This data table contains a set of admissions, that are assumed to be relevant for the classification task.

Column name	Description
patient_id	An identifier for patients across data tables
admission_id	A unique identifier for admissions
start_datetime	The start datetime of admission
end_datetime	The end datetime of admission

2. incidents This data table contains a set of incidents of patients included in the admissions dataset

Column	Description
patient_id	An identifier for patients across data tables
incident_id	A unique identifier for incidents
datetime	The datetime of the incident

3. notes This data table contains a set of clinical notes of patients included in the admissions dataset

Column	Description
patient_id	An identifier for patients across data tables
note_id	A unique identifier for notes
datetime	The datetime of the clinical note
text	The raw text of the clinical note

4. notes_full This data table consists of a large set of clinical notes, not necessarily included in the admissions dataset, used for training a paragraph2vec model for representing clinical notes.

Column	Description
text	The raw text of the clinical note
label	A label for the unsupervised learning task, can be arbitrary