Vulnerabilities of feature clustering in EIT radiomics

We aimed to determine whether unsupervised machine learning was able to discover latent and possibly clinically-relevant clusters, hidden in dynamic electrical impedance tomography (EIT) images within a population of mechanically ventilated COVID-19 patients. Dynamic EIT images visualize the distribution of electrical impedance within the patient's lungs, and is clinically used for the optimization of positive end-expiratory pressure (PEEP) when supporting patients with mechanical ventilation helping them to overcome their critical disease. The task selected here was to use unsupervised learning on 2D radiomics features extracted from clinical time-varying EIT images. The hypothesis was that we could find clusters of patients with similarly-appearing pathophysiology according to dynamic EIT. We were motivated to learn on which image-characteristics the resulting clusters were based and whether radiomic features confounded the results.

The setting of our study was an intensive care unit of a large tertiary academic hospital in the Netherlands. We used an unsupervised learning workflow that extracted 108 radiomics features and used principal components analysis, t-distributed stochastic neighbor embedding, and k-means to find latent clusters of similar dynamic lung impedance during the EIT-guided trials optimizing PEEP. Latent clusters were described by patient characteristics, and ventilation variables. We examined image acquisition parameters to control for confounding of radiomic features.

We included mechanically ventilated COVID-19 patients subjected to a clinical EIT measurement from March 2020 until September 2021, totalling 172 patients with 528 unique EIT-guided PEEP trials. We analyzed approximately 250,000 EIT images. Five latent clusters were identified each consisting of similar EIT-guided PEEP trials, which differed regarding patient characteristics, ventilation variables, chosen PEEP level, and clinical outcomes. Analysis showed that these clusters contained artifacts that encoded information about image acquisition process (the number of PEEP steps examined during an EIT-guided PEEP trial) rather than patient- or clinical characteristics.

Despite the clinically significant differences observed in the clusters identified by unsupervised machine learning, our analysis showed that the clusters were ultimately confounded by the number of PEEP steps set by our ventilation practitioners at the bedside during the data acquisition process. To overcome this confounder effect, we possibly need more standardized acquisition of EIT-guided PEEP trials or a larger dataset than we currently have. Nevertheless, this clustering may still hold clinical relevance, as it shows that the EIT-guided PEEP trial is applied differently across patients, and these differences are associated with distinct clinical outcomes. This, in turn, implies that the ventilation practitioner-consciously or unconsciously-makes decisions during the process that ultimately have a strong association with patient outcomes.