A complexity evaluation method based on interpretable deep learning models is proposed to assess the complexity of Human-Machine Interfaces (HMI) in automotive intelligent cockpits. This method aims to optimize the HMI design in intelligent cockpits, enhancing driver attention while allowing for evaluation during low-fidelity design stages or in no-code development environments, thus reducing design time and personnel costs. The evaluation process includes both static HMI image analysis and post-software development assessments. It involves forward model inference, selecting relevant layers, calculating gradients, and generating heatmaps to provide a comprehensive evaluation and visual feedback of the interface. Examining the heatmaps, we analyze whether the model focuses on the correct regions, detecting biases or misunderstandings. This helps understand model behavior, improve model performance, and enhance interpretability. Results show that the generated heatmaps effectively predict webpage complexity and are closely related to user-perceived complexity. Ultimately, comprehensive evaluations using the NASA-TLX scale validate the method’s effectiveness in reducing driver cognitive load and improving driving safety, significantly enhancing user experience and aiding developers in optimizing HMI interfaces.