The BAE project: Using AI models to predict patients who are likely to have Multi Drug Resistant

Dr Ke Yuhe, Anaesthesiology Resident, shares on her recent experience competing at the NUS Healthcare Datathon 2020.

---

The NUS Healthcare Datathon 2020 is an event co organised by NUS, NUHS and MIT Critcal Data. We had over 50 international teams formed from Singapore, China, Taiwan, Australia, Japan, South Korea, and even Russia. The teams consisted of clinicians, data scientists and innovators in healthcare to address current problems in healthcare with data analytics technologies. The event strengthens cross-disciplinary collaboration around secondary analysis of electronic health record data, helping to pave the way for reliable knowledge to be translated into action, for the benefit of patients.

Our team members consist of, from the clinical side, Dr Hairil, Yuhe and Nicholas Shannon. And from the data scientist side, Kien from NUS SSHSPH, and other analysts joining from America and South Africa.

Our team looked at the series of patients from the ICU-MIMIC IV dataset, where, for the first time in MIMIC, we were able to access the pre-ICU data. We performed feature extractions on parameters such as pre-morbidities, recent ICU stay and hospital admissions, recent medications (antibiotics, immunosuppressants, anti-oncological agents), and patient parameters in their first 6 hours of ICU stay. The groups we focused on were patients who grew MRDOs from cultures taken in the first 3 days of their ICU stay, and a comparison group of patients who had a positive culture growth for similar organisms, but which were sensitive to standard antibiotics. The clinical features for each group were run through several machine learning models to predict the growth of MDRO from data available on arrival and in the first 6 hours of ICU stay. Several AI models were developed to predict the factors which predispose a patient to have multidrug-resistant organisms (MDRO) upon admission to ICU. This project provided proof of principle for the creation of a predictive scoring system to identify patients who are more likely to grow MDROs and hence would be candidates for early escalation of antibiotic therapy

We were thrilled to have won the 1st runner up with the project, and only lost the crowd favourite by 1% of the vote! Nonetheless, this highlights the importance of the subject matter. We hope to extend the project to real-life applications with our own hospital data, and we envision a future of individualized antibiotics for every patient.