Introduction Intracranial hypertension (IH) is a harbinger of secondary brain injury in patients suffering from traumatic brain injury (TBI), can be mitigated at the Intensive Care Unit (ICU) and is associated with a poor prognosis. Current clinical practice consists of treating IH once it has occurred, by medical or surgical interventions. This is later than desired, as secondary injury has already been initiated. A pre-emptive approach may be preferable, and seems possible since many physiological variables that may aggravate IH are known and can be managed clinically. The aim of this research is to develop a machine learning method that is able to predict whether or not a patient will develop IH in the near future during ICU stay. Methods A cohort of 114 patients with TBI admitted to the ICU of Erasmus MC was selected. Long Short Term Memory (LSTM) models were trained and evaluated with 26 clinical variables to predict IH. The effect of the length of the minimal IH period, the length of the prediction window and the number of included variables was evaluated. Primary outcome measures were the model loss, accuracy, and Area Under the receiver operating characteristic Curve (AUC). Results We achieved a mean AUC of 0,83 [95% CI: 0,68-0,98] with a model predicting periods of ICP≥20mmHg lasting at least 15 minutes, using a prediction window of 30 minutes and using only the ICP and mean arterial blood pressure (MAP). All models showed decreasing training and validation loss values during the first few epochs of model training. Thereafter, the training loss continued to decrease while the validation loss started to increase. Conclusion We developed a LSTM model that was able to predict, with a mean AUC of 0.83 [95% CI: 0,68-0,98], the occurrence of IH after half an hour based on the ICP and MAP. Adding more clinical variables resulted in overtrained models. 

Background: Undetected Intellectual Disability (ID) can lead to chronic stress due to overestimation by society. Chronic stress can cause stress-related health issues, like hypertension, chronic fatigue and abdominal complaints. When a physician (General Practitioner (GP) or medical specialist) does not recognize that a patient has ID, the relation with stress may go unnoticed. In that case, the complaint is often treated as a purely somatic problem, while the underlying cause (overestimation due to unrecognized ID) remains untreated. This can increase healthcare consumption and impair the patient’s quality of life. While physicians with ID-expertise can recognize subtle signs of mild ID, physicians without extensive experience will easily overlook the ID. To improve medical care for patients with ID, we aim to improve ID detection among physicians. As it is not feasible to give all individual doctors an ‘ID-recognition training’, we study the possibility of using AI to improve ID detection. In the past years, we have been working on an ‘ID Alert’ (IDA) using ML. In previous phases of the IDA project, structured Electronic Health Record (EHR) data was used for the creation of an IDA. In addition, in the current study, we investigate the use of unstructured EHR data (clinical text). Methods: We analyzed unstructured correspondence files of 200 ID-adults and 200 non-ID adults of Novicare, an organization that provides multidisciplinary care to clients with complex and chronic conditions in intra- and extramural settings. Structured clinical data was unavailable. Therefore, we used an automated method of text extraction, de-identification and two types of feature extraction (bag-of-words and clinical concept extraction). Features were compared between ID-adults and non-ID adults. Significant features that were unlikely to be intrinsically different between ID- and non-ID adults were excluded. The remaining significant features were used for the training and evaluation (10-fold stratified cross-validation) of two Gradient Boosting Classifiers. Results: Most features differed significantly between ID- and non-ID adults due to confounders such as differences in age, type of care and the doctor’s word choice (which is inherent to the specialty and training of the doctor). Significant ‘unbiased’ features identified by both types of feature extraction methods are epilepsy, emotional disturbance (tension, arousal, agitation), visual or hearing problems and the presence of family members during consult. The developed ML models showed Areas Under the Curve (AUCs) of 0.98 and 0.89 for bag-of-words and clinical concepts, respectively. Conclusion: This is the first study to investigate the use of unstructured correspondence files for developing an IDA. The developed models show a very high performance. Despite efforts to mitigate the effect of confounders, limitations may have influenced generalizability. Therefore, external validation of the proposed methods is necessary in future research.

Myeloproliferative Neoplasms (MPNs) are a group of bone marrow diseases with potentially lethal cardio-vascular complications. Two sub-diseases of MPN are Essential Thrombocytosis (ET) and Polycythemia Vera (PV), which are recognised by an abnormal blood count of respectively thrombocytes and red blood cells. If an MPN is treated appropriately, complications for patients are reduced, leading to a relative increase of patients life expectancy. However, MPN is often recognised long after the first clinical signs. 1/4 of MPN patients already had abnormal blood measurements for longer than 1 year in advance of their diagnosis. Therefore, there is the call for methods for earlier recognition of MPN. Screening like methods could be useful to alert clinicians in case of a suspected case. Although genetic testing is conclusive in recognising MPN, high costs make that they are only applied in case of already clinically suspected MPN. In this thesis, the outlines of a method are proposed for early detection of MPN patients based on blood measurements in the general hospital laboratory workflow. A two stage solution is proposed: • Stage 1: Filter on regular blood measurements (combined with demographic data); • Stage 2: Filter based on microscopy imaging of blood. The primary scope of this thesis is the development of the first stage for ET and PV subtypes of MPN. A machine learning algorithm called XGBoost is utilized to develop classification algorithms for ET and PV in this stage. Patients with elevated blood platelet counts (ET marker) or elevated red blood cell indicators (PV marker) were separately included in a nested cross validation setup for training and testing of the algorithms. For ET vs control classification, mean metrics obtained during cross validation are AUC: 0.87, recall (sensitivity): 0.74 and specificity: 0.84. For PV vs control corresponding metrics are respectively 0.86, 0.66 and 0.87. Regarding the development of methods for stage 2, a first step is set. A XGBoost model using cell counts from microscopy images as features results in mean AUC, recall and specificity scores of 0.67, 0.78 and 0.80 respectively when trained and tested using nested cross validation. Training a Convolutional Neural Network (CNN) to take microscopy images as input and return MPN vs control classification resulted in an algorithm which only predicted control cases. These results give an indication of the potential of microscopy for automated MPN recognition, calling for further development of the stage 2 filter. With this proposed laboratory population screening method and the developed blood measurement based filtering, a next step is set toward early detection of MPN in order to prevent (lethal) MPN related complications.

Introduction: Magnetic Resonance Imaging is a commonly used technique for the initial diagnosis of gliomas. T1, T2, T2-FLAIR, and post-contrast T1 with gadolinium-based contrast agents (GBCAs) can show tumor characteristics. However, using this contrast agent poses a risk to patients with kidney failures, has environmental impact, and increases cost. To address these issues, we aimed to evaluate the potential of deep learning in generating synthetic post-contrast T1 images without using contrast agents. Method: The project investigates the potential of using deep learning (DL) to generate synthetic post-contrast T1 images based on T1, T2, and T2-FLAIR provided by the Erasmus Glioma Database. Exploring different model architectures, loss functions, and input sizes to discover the optimal approach. Results: Results show that individual loss functions, input size, and model complexity slightly impact the accuracy of synthetic post-contrast T1 images. Combining loss functions, however, was the most promising approach for image generation. Models trained with ℒm could generate low detail enhancement. Resulting in 0.0478±0.0076, 0.0139±0.0036, and 0.879±0.024 for MAE, MSE, and SSIM, respectively. Conclusion: The study's findings indicate that DL is promising for generating synthetic post-contrast T1 images without using GBCAs. However, further research is required to generate realistic synthetic post-contrast T1 images. The study, however, provides a basis for future work and highlights the importance of reducing the use of GBCAs in clinical practice.

Gastrointestinal stromal tumours (GISTs) are rare mesenchymal neoplasms with a worldwide incidence of one or two per 100,000. The tumours affect the entire gastrointestinal tract, but most commonly the stomach and small intestine. Neoadjuvant tyrosine kinase inhibitors (TKI) are administered in a selection of GIST patients to attain size reduction of the primary tumour and improve chances of complete resection. Response monitoring in GISTs is complex due to the presence of intra- and intertumoral heterogeneity and lack of pathological criteria to define response. Nonetheless, it is of importance to evaluate the efficacy of TKI treatment at an early stage in order to optimise treatment. In particular, early cessation of ineffective treatment is of importance in these patients, preventing unnecessary side-effects and healthcare costs. Medical imaging plays an important role to non-invasively predict and monitor treatment response of GIST patients undergoing TKI-treatment. This thesis aims to improve understanding of contrast-enhanced computed tomography (CE-CT) and 2-deoxy-2-[18F]fluoro-D-glucose ([18F]FDG) PET/CT imaging parameters to allow prediction and monitoring of neoadjuvant treatment response in GIST patients. In Chapter 1, the added value of CE-CT and [18F]FDG-PET imaging for early prediction and monitoring of treatment response in GISTs is investigated by means of a systematic literature review. Results of this study show that heterogeneous enhancement patterns on baseline CE-CT imaging were considered to be predictive for high-risk GISTs, reflecting neovascularisation and the presence of necrosis. Current CE-CT radiographic response criteria (i.e., RECIST 1.1 and Choi) are still lacking sensitivity and are prone to errors when predicting or monitoring treatment response. Metabolic changes on [18F]FDG-PET imaging seem to precede morphological changes in size in GIST lesions and were more strongly correlated with tumour response. Although CE-CT and [18F]FDG-PET can aid in the prediction and monitoring in GIST patients, further research on cost-effectiveness is recommended. Chapter 2 evaluates the efficacy of current radiological response criteria (RECIST 1.1, Choi and tumour volumetry) in predicting response to neoadjuvant systemic therapy, by comparing radiological response criteria with the achieved surgical benefit. Results show that size-based criteria (RECIST 1.1 and volumetry) accurately reflect surgical benefit in GIST patients treated with neoadjuvant systemic therapy (accuracy of 76.3% and 86.6% respectively) and are less prone to scanner and imaging protocol variabilities, when compared to the Choi criteria (68.4%). In addition to volumetry, quantitative radiomic models using CE-CT and [18F]FDG-PET imaging features were trained to predict response at baseline. Preliminary results of this study are described in Chapter 3. The radiomic models presented in this study generally had a poor performance and can therefore not yet be applied in a clinical setting. To improve performance and generalisability, future research should focus on a bigger patient population and harmonisation of acquisition protocols. The conclusion of this thesis supports the utilisation of tumour diameters for radiological response assessment, where RECIST 1.1 response criteria had an accuracy of 80.0% to correctly predict volumetric response after the first response follow-up CE-CT scan. When properly executed, these manual measurements could aid in early surgical decision making.