Risk-Aware Decision Making

Risk-Aware Decision Making for Service Robots

to Minimize Risk of Patient Falls in Hospitals

Roya Sabbagh Novin, Amir Yazdani, Andrew Merryweather, Tucker Hermans

University of Utah, Utah Robotics Center

Abstract

Planning under uncertainty is a crucial capability for autonomous systems to operate reliably in uncertain and dynamic environments. The concern of patient safety becomes even more critical in healthcare settings where robots interact with humans. In this paper, we propose a novel risk-aware planning framework to minimize the risk of patient falls by providing a patient with an assistive device. Our approach combines learning-based prediction with model-based control to plan for the fall prevention tasks. This provides advantages compared to end-to-end learning methods in which the robot's performance is limited to specific scenarios, or purely model-based approaches that use relatively simple function approximators and are prone to high modeling errors. We compare two different risk metrics and the combination of them and the results from various simulated scenarios show that using the proposed cost function, the robot can plan interventions to avoid high fall score events.

[Main Paper + Supplementary PDF]

Source Code

Github repo: risk-aware decision making

Video

Overall Framework

Fall Score Distributions

The following figures provide the fall score distributions from various experiments initiated from different locations within the hospital room. In each figure, we compare the results using different methods and cost functions. First, comparing the “no-intervention” and “deterministic” plots, we see the effect of utilizing a service robot to assist a hospital patient. Next, we can observe that using a probabilistic method has improved the overall performance of our planning considering a broad range of predictions. Finally, we can see that adding the CVaR risk metric has reduced the number of events with high fall score which is the main purpose of this research.

Fig. 1. Fall distribution from 20 scenarios in which the patient’s initial pose is near the right side of the bed.

Fig. 2. Fall distribution from 20 scenarios in which the patient’s initial pose is near the toilet.

Fig. 3. Fall distribution from 20 scenarios in which the patient’s initial pose is near the left side of the bed.

Fig. 4. Fall distribution from 20 scenarios in which the patient’s initial pose is near the visitor chair.

For each initial pose, we run 20 scenarios with random intentions (sampled from a prior distribution over the intention set) and calculate the fall score along the actual patient path, before and after intervention. This results in a distribution of fall scores for a single initial pose which we show in the figure below. We provide the distributions for five different methods. We can see the effect of CVaR and expected cost functions on the mean and the area under the tail of the distribution. Using our probabilistic approach, we can reduce the number of rare events with high fall score. We also see that the "CVaR" and "expected + CVaR" cost functions perform best among all methods. Although the robot is successful in recognizing high fall score events even with low probability and takes action accordingly, the amount of change it must make was not drastic due to the limited assistance of the walker. In our future studies, we would like to explore other assistance options.

Fig. 5: Fall distribution from 20 scenarios in which the patient’s initial pose is near the right side of the bed.

The following figures provides the mean and CVaR of the fall score distributions from five methods for various initial poses. Here, we see that for some scenarios, different methods can result in similar performances. This is mainly in scenarios where the robot can provide the walker early in the patient's trajectory (such as scenarios where the patient starts from the visitor chair), resulting in the best performance possible for that scenario. However, there are some scenarios in which the robot must decide between multiple choices (such as the scenarios with the patient initial pose near the right side of the bed), creating the main challenge discussed in our problem. We see that, in those scenarios, our proposed cost function outperforms the other methods.

Fig. 6: Mean and CVaR of fall scores distributions from 20 sce narios for four initial patient poses using different fall-prevention approaches. Overall, “Expected+CVaR Cost” approach has a betterperformance considering these two metrics.

Page updated

Google Sites

Report abuse