Kinetic Instrumentation of Sit-to-Walk Component of Timed Up and Go Test

Functional mobility assessment is crucial for identifying fall risk and guiding rehabilitation in older adults. The Timed Up and Go (TUG) test is a widely used clinical measure of mobility, but its primary metric, total time to complete, only offers a basic overview of the complex movement process. Wearable inertial measurement units (IMUs) used in instrumented versions of the TUG provide valuable kinematic and spatiotemporal data. However, kinetic variables such as ground reaction forces (GRF), ground reaction moments (GRM), and center of pressure (CoP) are mostly measurable only in laboratory settings because they require fixed force plates. These kinetic measures provide important clinical insights into lower-limb loading, postural control, and balance strategies that kinematics alone cannot reveal. Therefore, developing methods to estimate kinetic variables from wearable sensors during the TUG could enhance the clinical usefulness of this assessment without requiring specialized lab equipment. This thesis aims to develop and evaluate machine learning techniques to estimate these kinetic parameters during the initial transition of the TUG, the sit-to-walk (STW).

A PRISMA-compliant systematic review was conducted to outline the current landscape of machine learning and deep learning methods for GRF estimation from IMU data. The review included 22 studies, examining participant groups, sensor setups, preprocessing methods, model types, and evaluation approaches. Although individual studies compared selected architectures, differences in normalization, validation protocols and performance metrics limited meaningful comparisons across studies.

To address these gaps, an experimental study tested 13 model configurations to predict continuous three-dimensional GRF waveforms during STW from a single sacrum-mounted IMU, using a publicly available dataset of 65 participants. All models were evaluated with a consistent pipeline. Results showed that no single architecture dominated across all three GRF axes. However, the DS-Conformer was the most consistently high-performing, achieving the best vertical GRF prediction (RMSE 4.05 %BW, R2 = 0.985) and the highest waveform agreement on the anteroposterior and mediolateral axes.

A follow-up study expanded the DS-Conformer to estimate all eight kinetic components during STW: three GRF, three GRM, and two CoP channels. Both experimental studies used the same dataset. Testing on held-out participants and age groups supports prediction for new people within that dataset, but does not establish performance with a different sensor, laboratory or acquisition setup. Leave-one-out cross-validation across 65 subjects showed normalized RMSEs from 3.78% to 11.48% and Pearson correlations between 0.787 and 0.990. Predictions for GRF and GRM were consistent across ages, but CoP accuracy varied significantly between age groups in the mediolateral direction, highlighting the limited ability of a single midline sensor to detect inter-limb lateral weight-shifting strategies that differ among individuals and age groups.

This thesis demonstrates the potential for wearable kinetic assessment during the STW transition of the TUG test using a single body-worn IMU and deep learning. The proposed framework is a step towards providing mechanistic kinetic information during routine TUG assessments that time alone cannot, without requiring laboratory infrastructure. However, several questions remain open. Validation in clinical populations with asymmetric or impaired movement has not yet been undertaken, and normative reference ranges and outcome-linked thresholds for STW kinetics must be established before predicted waveforms can inform clinical decisions. The mediolateral bottleneck also warrants targeted work on sensor expansion or on the fusion of sensing modalities to recover lateral balance information that a single midline sensor cannot capture. The framework currently covers only the first transition of TUG. Future work should extend the framework to the remaining TUG subtasks.