Driver fatigue is the act of operating a motor vehicle while drowsy, which poses a significant risk to the individual and everyone around them. In the UK, 10–20% of crashes are attributed to fatigue. To address this issue, a novel multi-level, multi-feature fusion approach is proposed, achieving 98.5% accuracy in fatigue detection for a 3-second video clip and diagnosing specific fatigue states (microsleeping or yawning) with 100% accuracy. This approach combines spatial information extracted via MediaPipe with a dual LSTM framework, providing an advantage over single-model systems by enabling both detection and diagnosis of fatigue. By analyzing features such as eye, mouth, and head pose, the system ensures a robust and accurate prediction. The NITYMED dataset was used for training and testing, with videos segmented into 3-second clips. Additionally, a new dataset was collected and segmented for validation, demonstrating model generalizability. Latency analysis showed mean inference times of 2.9 ms per LSTM model, reinforcing the system’s suitability for real-time applications. Potential future improvements include refining model robustness by evaluating its performance across individuals of diverse ethnic backgrounds and clip duration experimentation. An effective implementation of this system has the potential to dramatically reduce fatigue-induced driving accidents, contributing to overall road safety.