Presentation (oral) at CAMO-Net AI meeting in London.

This presentation explores the transformative potential of leveraging time series dependencies through advanced machine learning techniques in the clinical management of acute febrile illnesses. It explores the use of machine learning models, such as Recurrent Neural Networks (RNN) and Long Short-Term Memory (LSTM), to predict bloodstream infections based on temporal data and highlights the potential of real-time monitoring systems that enhance early detection, diagnosis, and treatment strategies. To conclude, it discusses a study using photoplethysmogram (PPG) signals to predict Dengue patient outcomes and risk, utilizing a Convolutional Neural Network (CNN) and multimodal data fusion. The presentation argues for the importance of considering time series dependencies and the increasing availability of data when developing machine learning models to facilitate personalized and timely interventions, ultimately contributing to more effective healthcare outcomes.

This presentation discusses the challenges and limitations of diagnosing severe bacterial infections, specifically sepsis, using data-driven approaches and emphasizes the importance of considering past medical history and co-morbidities, as well as continuous measurement and re-evaluation of diagnoses. It also highlights the need to define a consistent reference standard for serious bacterial infections and the potential use of machine learning diagnostics from various data streams, including electronic health records. To conclude, it acknowledges the complexities and unanswered questions related to the application, implementation, interpretability, and complexity of developing a model for accurately predicting bacterial infections.

This presentation discusses the challenges of diagnosing and providing support for Dengue disease, a mosquito-borne illness, in tropical and subtropical regions where data may be limited and emphasizes the importance of efficient resource allocation. An unsupervised approach based on case-based reasoning combined with a 2D embedding space representation is proposed to identify similar cases and learn from their outcomes. Various techniques are explored to generate the embbedings, including PCA, T-SNE, UMAP, and self-organizing maps, ultimately settling on an Autoencoder algorithm. The latent space produced is thoroughly explained demonstratring that the results align with established clinical characteristics exhibited by patients with dengue infection, as well as features of disease progression. Furthermore, similar clinical phenotypes are represented close to each other in the latent space and clustered according to outcomes broadly described by the World Health Organisation dengue guidelines. To conclude, these findings are included into a clinical decision support system to support monitoring and management of dengue and other febrile illness.

The presentation provides an update on simulation strategies including agent-based, discrete event simulation and other system dynamics. It specifically shows the use of agent-based modeling for social unrest simulation and discrete event simulation for hospital processes. It also provides a demo of a discrete event simulation system of a hospital where patients are registered as events with specific event times, executed when the time counter reached the event times, and removed once completed. Overall, these simulation methods offer valuable insights into complex systems like hospitals, social unrest, and other applications.