Multimodal 100 day postoperative mortality prediction for geriatric patients with a hip fracture: working towards clinical implementation

Hip fractures are a highly prevalent type of fracture among elderly patients with a high chance of mortality (13% within a 100 days post surgery), high healthcare costs and a heavy rehabilitation process. The ability to forecast mortality can aid making a more informed treatment plan, such as palliative treatment for highly frail patients. This thesis builds upon an existing multimodal model developed by ZGT (Hospitalgroup Twente) to predict 30-day and 100-day post surgery mortality for elderly patients (age >70). A baseline model is compared with a new iteration that includes new data and additional features. Such features inlcude NLP extracted comorbidities and a similarity score with patients who received palliative treatment. A Key finding is that changing the target outcome to 100-day mortality significantly improves all performance metrics. The inclusion of comorbidity data shows mixed results; it does not improve the original model. However, adding comorbidities to a model with static data increases the AUCPR by 2 percentage points. Furthermore, the integration of similarity scores from NOM patients demonstrates a substantial improvement in Maximum precision, with a 20-percentage-point increase, an enhancement of AUCPR by 6 percentage points. The AUCROC remains the same at 0.77.