PhD Defence Sjoerd Garssen | Exploring supportive signals: Machine learning and wearable sensors for patient monitoring

Exploring supportive signals: Machine learning and wearable sensors for patient monitoring

Sjoerd Garssen is a PhD student in the Department of Health Technology & Services Research. (Co)Promotors are prof.dr. C.J.M. Doggen, prof.dr.ir. B.P. Veldkamp, dr. M. Amir Haeri and dr. S.F. oude Wesselink from the Faculty of Behavioural, Management and Social Sciences.

Healthcare must become more efficient to meet its increasing demand, which also concerns Acute Medical Units (AMUs). In AMUs, patients undergo observation, receive initial treatment, and have their vital signs monitored intermittently. Wearable sensors monitor vital signs continuously and may support discharge decisions in AMUs. Machine learning may also support these decisions. Besides, wearable sensors and machine learning may be used for remote patient monitoring, thereby contributing to the transformation of care from hospitals to home. For remotely monitored patients, early detection of clinical deterioration is crucial to ensure that care can be upscaled in time.

The aim of this thesis was to investigate whether machine learning and wearable sensors can support the discharge decisions in an AMU and detect signs of deterioration of discharged patients at home.

A randomized controlled trial was conducted. Included were AMU patients whose discharge destination, either home or a transfer to another ward, was uncertain at the moment of AMU admission. This trial showed similar proportions of patients who were safely discharged home directly from the AMU between control (30.8%) and sensor groups (33.2%, p=0.62). Length of stay, the number of Intensive Care Unit admissions, and the number of Rapid Response team calls were also similar.

The resulting dataset was used to explore machine learning for assessing physiological discharge fitness. Models only using available EMR data reached a mean AUROC of up to 0.69 for the control and 0.70 for the sensor group, which increased to 0.73 with additional wearable sensor data.

The same dataset and a dataset of postoperative patients were used to explore whether personalized anomaly detection models could detect early signs of patient deterioration at home. These models were able to detect early signs of patient deterioration at home in former AMU patients to a certain extent (AUROC up to 0.69), but not for postoperative patients (AUROC up to 0.44).

In conclusion, machine learning and wearable sensors may support the discharge decision making in AMUs and detect early signs of patient deterioration at home if used appropriately. However, current performance is not yet sufficient for clinical implementation.