Neurological conditions often manifest as gait disorders, frequently linked to spasticity. Botulinum Toxin Type A (BTX-A) injections commonly treat spasticity-related gait issues. Achieving optimal treatment outcomes with a favourable benefit-risk ratio remains crucial. This paper proposes an innovative method to predict post-BTX-A treatment knee and ankle kinematics by leveraging pre-treatment data and treatment specifics. Our approach relies on a bidirectional long-short-term memory (Bi-LSTM) network integrated with an attention mechanism within a deep learning architecture. The primary objective is to assess the efficacy of this methodology in accurately forecasting gait cycle kinematics for the knee and ankle joints after BTX-A intervention. Two deep learning models are designed, integrating categorical medical treatment data (MTD) representing injected muscles: (1) embedded within the Bi-LSTM network’s hidden layers and (2) through a gating mechanism. These architectures aim to model interactions among various treatment combinations when multiple muscles are injected simultaneously. Through comparative analysis with state-of-the-art approaches, our study demonstrates that incorporating attention mechanisms yields superior results. The average root-mean-squared error for predictions stands at 3.03 \(^{\circ }\) ( \(R^2\) = 0.87) for knee kinematics and 2.18 \(^{\circ }\) ( \(R^2\) = 0.83) for ankle kinematics. Our findings conclusively indicate that our proposed approach surpasses existing methods, offering higher predictive accuracy for post-BTX-A treatment kinematics.