Towards Personalized Medicine: Using Retrospective Data to Predict the Anticoagulation Effect of Heparin

Abstract

Heparin is an essential drug in anticoagulant therapy and used to prolong the bloods coagulation time, thus being useful in many clinical scenarios. However, according to it's highly variable dose-response relationship, heparin is frequently misdosed which goes along with an increased risk of severe side effects and 30-day mortality for some indications. A statistic model that considers individual patient factors to predict the anticoagulation effect of heparin would therefore be of high clinical relevance. Accordingly, this work aims to develop such a multi-featured model to moreover question today's heparin dosing guidelines which are solely based on the patient's weight. Hence, data of about 5,686 heparin therapies has been extracted from MIMIC-IV, a publicly available database that contains retrospective medical data. Subsequently, the data has been used to train and evaluate different regression models that were provided by scikit-learn, a python library that integrates a wide range of state-of-the-art machine learning algorithms. Using the R2 metric to quantify the prediction performance, models that only consider patientweight were compared to models that additionally consider variables such as age, gender, ethnicity etc.\ which are currently not considered when deciding about a patients heparin dosage. It has been found that a polynomial regression model is best suited to predict the anticoagulant outcome of heparin therapy, achieving R2 scores of 0.1850 and 0.2636 for the purely weight-based and multi-featured model respectively. This implies a recommendation to extend current heparin dosing guidelines. For example, it has specifically been found that patients with higher age and female gender are more sensitive with respect to heparin administration, thus requiring lower average dosages to achieve the same anticoagulant outcome when compared to young males. Altogether, this work demonstrates the potential of retrospective medical data to provide a more personalized medicine. The results that have been achieved in this thesis could been build on when developing decision support systems for today's clinical practice whilst initiating change processes regarding current heparin dosing guidelines in intensive care unit settings.

Data Extraction

First of, data from MIMIC-IV has been extracted according to the following workflow:

• Extraction of Heparin Data via hep_data.sql
• Select only Heparin Data of Cohort 1 via hep_data_cohort1.sql
• Preprocess the data with Python via cohort1_hep_data_preprocessed.ipynb
• Extraction of APTT Data via aptt_data.sql (Note that other Features have been added here as well, e.g. SOFA Score according to sofa.sql
• Select only APTT Data of Cohort 1 via aptt_data_cohort1.sql
• Merge Heparin and APTT data for the final dataframe via final_data.sql

Data Analysis

The data analysis has then been performed as following:

• The first step was to show some data plots, as can be seen in: Cohort1_Data_Visualization.ipynb
• Them, the idea behind the pharmacokinetic models have been illustrated according to: Visualization of Pharmacokinetic Models.ipynb
• However, the development of the actual third Pharmacokinetic Model is described in more detail here: Logic of Pharmacokinetic Model 3.ipynb
• The Pharmacokinetic Models have then been analysed in Evaluation of Pharmacokinetic Models.ipynb
• Accordingly, the Features have been evaluated in Evaluation of Feature Integration.ipynb
• Lastly, a logistic regression model have been trained in Logistic Regression.ipynb