Research Objectives: To develop an AI-empowered technology that automatically evaluates sitting-pivot-transfer performance based on body motion tracking and ambient object recognition with a low-cost time-of-flight sensor.
Design: The Cross Industry Standard Process for Data Mining (CRISP-DM) approach with the data collected from multi-site clinical trials was adopted.
Setting: The multi-site clinical trials were assessed based on the Transfer Assessment Instrument (TAI), which is a clinical tool to be used by therapists to evaluate quality of the wheelchair transfer.
Participants: Fifty-five full-time wheelchair users were asked to perform 5 times of sitting-pivot-transfers from their wheelchair to a level-height bench using their everyday transfer techniques while 2 trained sitting-and-mobility specialists scored the TAI and a time-of-flight sensor simultaneously recorded their body motion and ambient object data in the format of 3D coordinates.
Interventions: Dataset was prepared: combining the body motion and ambient object data in 3D format (features/attributes) with the TAI assessment results (classes/labels). With the dataset, we have applied different machine learning algorithms with a k(5)-fold cross-validation as a validation scheme, and compared the results and selected the best fitting model.
Main Outcome Measures: The factors considered essential for a measure of transfer construct were measured, including setup of the wheelchair with respect to the target surface, use of conservation techniques (e.g., alternating leading/trailing arm, using handgrips), and quality of the transfer.
Results: Overall, our developed models achieved 93.76% accuracy (95.11% precision and 95.18% recall) with the following details (item [name of the machine algorithm/accuracy/precision/recall]): wheelchair-distance-from-target-surface [Ensemble/98.46%/99.17%/98.35%]; angle-between-wheelchair-and-target-surface [Ensemble/97.04%/98.96%/96.94%]; over-rear-wheel [Fine-KNN/94.85%/94.62%/98.88];
feet-in-a-stable-position [Quadratic-SVM/95.47%/95.16%/98.33%];
scooting-forward [Ensemble/96.36%/97.89%/97.89%]; hand-in-a-stable-position [Ensemble/98.55%/99.56%/98.69%]; leading-handgrip [Ensemble/95.62%/97.25%/96.20%]; trailing-handgrip [Ensemble/91.27%/91.98%/93.13];
flight-control [Cubic-SVM/90.18%/89.60%/94.51%]; leaning-forward [Ensemble/88.36%/91.41%/91.41%]; head-hip-direction [Ensemble/89.82%/90.96%/89.88%]; leading-arm-position [Ensemble/98.18%/99.03%/98.55%]; and landing-control [Ensemble/84.73%/90.80%/84.57%]
Conclusions: This study demonstrates that a sitting-pivot-transfer can be successfully quantified and assessed using artificial intelligence. It holds good potential to relieve burden on the therapist by automating the process of transfer assessment and documentation of results, eliminating the need for extensive training and reducing the time taken to assess a patient.
Leave a Reply