Publications by authors named "L E M Roggeveen"
Article Synopsis
- Reinforcement learning (RL) has potential in intensive care medicine due to the wealth of data and real-time decision-making needs, but trust and safety issues hinder its implementation for clinical support, particularly in optimizing ventilator settings for COVID-19 patients.
- A novel method called cross off-policy evaluation (OPE) was developed to evaluate RL models, using a large dataset from Dutch ICUs and focusing on ventilator settings, with interim and final rewards based on gas exchange indices and patient outcomes.
- The study found that while many RL policies performed well initially, almost half were deemed suboptimal under more rigorous evaluation, highlighting the importance of detailed clinical policy inspection and restrictions to ensure safety in patient care.
Tests of human brain circuit function typically require fixed equipment in lab environments. We have developed a smartphone-based platform for neurometric testing. This platform, which uses AI models like computer vision, is optimized for at-home use and produces reproducible, robust results on a battery of tests, including eyeblink conditioning, prepulse inhibition of acoustic startle response, and startle habituation.
View Article and Find Full Text PDFArticle Synopsis
- The increasing use of AI in healthcare highlights the need for standardizing medical data from different electronic healthcare record (EHR) systems, as they often use varied naming conventions for the same concepts.
- The study proposes an augmented intelligence method to align these different terminologies by predicting accurate medical concepts from raw EHR data, utilizing machine learning models trained on manually mapped data from multiple hospitals.
- Results show that the initial model achieved a precision score of 0.744 and a recall score of 0.771 when applied to a large dataset, indicating promising effectiveness in concept mapping across diverse EHR systems.
Article Synopsis
- Prone positioning has emerged as a key treatment for mechanically ventilated COVID-19 patients, but it requires significant labor and can have negative effects, making it essential to identify which patients will benefit from the procedure.
- A study utilized data from over 1,100 intubated patients across 25 hospitals in the Netherlands, applying various machine learning models to predict the success of prone positioning after 4 hours by evaluating improvements in respiratory metrics.
- Despite extensive analysis using different machine learning techniques, the study found that distinguishing between patients who would respond positively to prone positioning and those who would not had limited success, indicating potential challenges in predicting treatment outcomes.
Article Synopsis
- A study called AutoKinetics was developed to enhance antibiotic dosing in critically ill patients, addressing challenges in drug absorption and distribution.
- The trial involved 252 patients with sepsis or septic shock, comparing AutoKinetics personalized dosing against standard methods for four antibiotics.
- Results showed that AutoKinetics significantly improved dosing success for ciprofloxacin, with no increased risk of mortality or kidney issues, while also being safe and feasible.*