KIT2
Decision support for pre-hospital emergency care
KIT2 built an AI-driven assistant for the German TeleNotArzt (telemedical emergency physician) service, where clinicians must make many critical decisions in very little time. The system takes a patient's heterogeneous record — time-series vital signs, tabular clinical fields, and unstructured free-text notes — and returns probabilistic rankings of likely diagnoses, suitable treatments, and the physiological measurements that should be taken next.
- Deep multimodal learning architecture over time-series, tabular, and text inputs
- Missing-value imputation for a real (non-public) emergency dataset — denoising LSTM autoencoders outperformed classical methods
- Synthetic tabular data generation (WLSTM-GAN) to address privacy limits and data scarcity
- Evaluated with recall@k, macro/weighted AUC, and coverage-ratio metrics
