(G-252) Predicting delirium using ambient light and noise information in Intensive Care Units

Introduction:: Intensive Care Unit (ICU) environments are typically characterized by artificial light, noise from machinery, staff conversations and absence of windows[1-2]. This constant exposure to artificial light and noisescapes and the concomitant deprivation from natural light cycles can cause circadian desynchrony and delirium[3]. Delirium is an acute change in awareness and attention that develops over a short period of time and can be associated with memory deficit, disorientation, and hallucination[4]. Delirium significantly increases the morbidity and mortality in ICU patients[5]. Despite previous studies indicating significant differences in light and noise levels in ICU rooms hosting delirious patients[3] no studies have attempted to predict the onset of delirium solely based on these environmental factors. In this study we predicted the incidence of delirium from the beginning of a patient's ICU stay till 1 week after discharge using temporal sequences of ambient light and noise information during daytime and nighttime. We used sequential deep learning models such as 1-D convolutional neural nets (CNN), long short term memory (LSTM) and gated recurrent units (GRU) in our analysis. Our results indicate that noise levels are stronger predictors of delirium onset than light levels. Furthermore, interpretability analysis revealed that maximum nighttime and minimum daytime noise levels were the strongest predictors of delirium.

Materials and Methods:: Data for this study was collected from ICU rooms in the University of Florida (UF) Health Shands Hospital. Noise data was collected using AudioTools application installed on an iPod. Light data was collected using ActiGraph GTX3+ sensors. For some patients, both light and noise information were collected simultaneously using the Thunderboard multi-array sensor module. Delirium states were ascertained using the Confusion Assessment Method for Intensive Care Units (CAM-ICU), Richmond Agitation and Sedation Scale (RASS) and Glasgow Coma Scale (GCS) scores. These were administered daily by ICU nurses. Noise data was collected for 100 patients over a total of 445 days. Light data was collected for 110 patients over a total of 794 days. Combined noise and light information was available for 49 patients and 391 days. Statistical analysis revealed that nighttime and daytime noise levels were significantly different from each other. Therefore, light and noise data were divided into daytime levels and nighttime levels. Temporal deep learning methods such as 1-D CNN, LSTM and GRU were trained using daily average daytime and nighttime values of only noise, only light or combined light and noise information. The objective of these models was to correctly classify whether patients experienced delirium during a period starting from the first day of data collection to 1 week after the end of their ICU stay. The maximum time period of ICU stay in our cohort was 7 days, so the maximum duration for which the classification task was performed was 14 days.

Results, Conclusions, and Discussions:: Only noise information was able to classify delirious patients from non-delirious participants with Area Under the Curve (AUC) = 0.83, sensitivity = 0.50 and specificity = 0.71. Only light information was able to classify delirious patients from non-delirious participants with AUC = 0.61, sensitivity = 0.56 and specificity = 0.67. Combined light and noise information was able to classify delirious patients from non-delirious participants with AUC = 0.64, sensitivity = 0.50 and specificity = 0.68. Therefore, the noise information alone was a better predictor of delirium than the light or the combined datasets. Shapeley Additive Explanations (SHAP) interpretation plots revealed that the minimum daytime noise, and maximum nighttime noise were the strongest predictors in the noise model. Higher values of maximum nighttime noise predicted delirium onset but higher values of minimum daytime noise predicted no delirium. In the model using only light information, nighttime light was a stronger factor than daytime light, but the directionality of these features were ambiguous.

Acknowledgements (Optional): :

References (Optional): : [1] Lee HJ, Bae E, Lee HY, Lee SM, Lee J: Association of natural light exposure and delirium according to the presence or absence of windows in the intensive care unit. Acute Crit Care 2021, 36(4):332-341.
[2] Simons KS, Verweij E, Lemmens PMC, Jelfs S, Park M, Spronk PE, Sonneveld JPC, Feijen HM, van der Steen MS, Kohlrausch AG et al: Noise in the intensive care unit and its influence on sleep quality: a multicenter observational study in Dutch intensive care units. Crit Care 2018, 22(1):250.
[3] Davoudi A, Malhotra KR, Shickel B, Siegel S, Williams S, Ruppert M, Bihorac E, Ozrazgat-Baslanti T, Tighe PJ, Bihorac A et al: Intelligent ICU for Autonomous Patient Monitoring Using Pervasive Sensing and Deep Learning. Sci Rep 2019, 9(1):8020.
[4] Ali M, Cascella M: ICU delirium. 2020.
[5] Maldonado JR: Acute Brain Failure: Pathophysiology, Diagnosis, Management, and Sequelae of Delirium. Crit Care Clin 2017, 33(3):461-519.