Treatment Outcome Prediction Using Multi-Task Learning: Application to Botulinum Toxin in Gait Rehabilitation

Abstract

We propose a framework for optimizing personalized treatment outcomes for patients with neurological diseases. A typical consequence of such diseases is gait disorders, partially explained by command and muscle tone problems associated with spasticity. Intramuscular injection of botulinum toxin type A is a common treatment for spasticity. According to the patient's profile, offering the optimal treatment combined with the highest possible benefit-risk ratio is important. For the prediction of knee and ankle kinematics after botulinum toxin type A (BTX-A) treatment, we propose: (1) a regression strategy based on a multi-task architecture composed of LSTM models; (2) to introduce medical treatment data (MTD) for context modeling; and (3) a gating mechanism to model treatment interaction more efficiently. The proposed models were compared with and without metadata describing treatments and with serial models. Multi-task learning (MTL) achieved the lowest root-mean-squared error (RMSE) (5.60°) for traumatic brain injury (TBI) patients on knee trajectories and the lowest RMSE (3.77°) for cerebral palsy (CP) patients on ankle trajectories, with only a difference of 5.60° between actual and predicted. Overall, the best RMSE ranged from 5.24° to 6.24° for the MTL models. To the best of our knowledge, this is the first time that MTL has been used for post-treatment gait trajectory prediction. The MTL models outperformed the serial models, particularly when introducing treatment metadata. The gating mechanism is efficient in modeling treatment interaction and improving trajectory prediction.

Keywords: botulinum toxin; clinical gait analysis; deep learning; gait rehabilitation; long short-term memory; multi-task learning.

Figure 1

Example of the outcome of BTX-A treatment on gait (a) before treatment (b) after BTX-A treatment.

Figure 2

Clinical gait analysis. Different types of sensors are used to conduct kinematic and kinetic analyses of locomotion in gait labs. These may include optoelectronic motion capture, force platforms, electromyography, and IMU sensors, among others.

Figure 3

Process of converting one trial to one normalized cycle.

Figure 4

LSTM unit. The gates, which decide which part of the information to pass on, are orange. Green is the update to the memory cell.

Figure 5

LSTM architectures (Model 1 and Model 2) proposed in this work: (a) without MTD; (b) with MTD.

Figure 6

First Bi-LSTM architecture (Model 3) proposed in this work without considering MTD.

Figure 7

Multi-task learning architectures with Bi-LSTM sub-models; (a) Model 4: processing MTD internally in each sub-model; (b) Model 5: incorporating MTD through a gating mechanism; (c) Model 6: processing MTD internally in each sub-model using the Conv layer; (d) Model 7: incorporating MTD through a gating mechanism using the Conv layer.

Figure 8

Comparison of the post-treatment gait trajectory of the knee and ankle joints in a patient diagnosed with CP. The first three models (a–c) are serial (Model 1, Model 2, and Model 3), and the fourth model (d) (Model 7) is an MTL model, which represents the prediction of the knee joint. The sixth and seventh models (e,f) are serial (Model 1 and Model 3), and the last two models (g,h) (Model 4 and Model 7) are MTL models, which represent the prediction of the ankle joint.

Figure 8

Comparison of the post-treatment gait trajectory of the knee and ankle joints in a patient diagnosed with CP. The first three models (a–c) are serial (Model 1, Model 2, and Model 3), and the fourth model (d) (Model 7) is an MTL model, which represents the prediction of the knee joint. The sixth and seventh models (e,f) are serial (Model 1 and Model 3), and the last two models (g,h) (Model 4 and Model 7) are MTL models, which represent the prediction of the ankle joint.

Figure 9

Comparison of the post-treatment gait trajectory of the knee and ankle joint in a patient diagnosed with MS. The first two models (a,b) are serial (Model 1 and Model 2), and the third and fourth models (c,d) are MTL models (Model 4 and Model 6), which represents the prediction of the knee joint. The fifth, sixth and seventh models (e–g) are serial (Model 1, Model 2, and Model 3), and the last model (h) is an MTL model (Model 4), which represents the prediction of the ankle joint.

Figure 9

Comparison of the post-treatment gait trajectory of the knee and ankle joint in a patient diagnosed with MS. The first two models (a,b) are serial (Model 1 and Model 2), and the third and fourth models (c,d) are MTL models (Model 4 and Model 6), which represents the prediction of the knee joint. The fifth, sixth and seventh models (e–g) are serial (Model 1, Model 2, and Model 3), and the last model (h) is an MTL model (Model 4), which represents the prediction of the ankle joint.