Maasstad Hospital is a member of the Santeon hospital group. The ambition of Santeon is to improve healthcare for patients. The project in this internship also aims to improve patients’ health, specifically patients in the Intensive Care Unit (ICU).
The treatment of respiratory insufficient patients in the ICU consists of High Flow Nasal Oxygen or Cannula (HFNO or HFNC), among others. There is a substantial uncertainty about the optimal duration of this HFNO therapy and the chance of failure of this therapy. Failure of HFNO therapy will often lead to the progression to mechanical ventilation with intubation. This thesis project researched parameters and predictive models to choose the appropriate treatment, meaning continuing HFNO therapy or escalation to mechanical ventilation by intubating the patient.
In this thesis project, the goal was to develop a Machine Learning (ML) model that can predict intubation at a certain point in time and thereby show that HFNO therapy will not be sufficient. With this eventual model, it would be possible to determine if intubation is necessary on the first day of ICU admission. The proposed model could lead to more elective or early intubation.
The intended ML model was achieved with a two-sided method. Firstly, an aggregated data set was used to compute three different models. These were two tree-based models, a Random Forest (RF) and a Gradient Boosting Model (GBM), and a Logistic Regression Model (LRM). The other parallel method made use of a data set with repeated measurements of vital parameters, such as heart rate. This method resulted in a so-called joint model, which is a combination of Linear Mixed Effects Models and in the second step also an RF, GBM, and LRM.
A nested cross-validation was implemented to test the above-described models, three feature selection methods, and three scaling methods. From the nested cross-validation, the best-performing model was found and tested in the evaluation.
For the evaluation of the models an extra data set was used. This external data set was retrieved via a data request to the Santeon ’Beheercommissie’ or data management committee. This data set did contain repeated measurements, but not enough to validate the joint model. Therefore, only the models developed with the aggregated data set could be externally validated.
The two-sided method resulted in an RF with no feature selection and no scaling having the best performance using the aggregated data set, namely an AUC of 0.694 (standard deviation 0.05). The joint model resulted in an RF with no feature selection and Power Transformer scaling with the best performance. It had a performance value of 0.681 (standard deviation 0.07). The external validation of the aggregated data model resulted in an AUC of 0.559. The internal validation of the joint model gave an AUC of 0.699. The precision, recall and f1-score showed that all the models performed better for class 0: the non-intubated patients.
The best-performing aggregated data model shows potential and proves that it is possible to predict intubation using AI, but in its current state is far from implementation. It is therefore advised to train the model with a larger training data set that contains multiple hospitals and perform an external validation
with a validation data set that meets the requirements.