This study presents a safety-prioritized deep reinforcement learning framework for navigation and safe lane-change decision-making in autonomous driving on highways. The framework introduces a discrete action space that combines streamlined lateral lane-change actions with longitudinal acceleration and deceleration maneuvers. A key feature of the framework is the integration of probabilistic risk assessment, grounded in Bayesian inference, to account for uncertainties in the behavior of surrounding vehicles. Unlike many prior studies, this work implements a fully autonomous decision-making system that models both lateral and longitudinal actions within a unified structure. Two deep reinforcement learning algorithms, deep Q-network and proximal policy optimization, are employed to enable the agent to make risk-aware decisions in dynamic environments. A multi-objective reward function is designed that prioritizes safety over efficiency while also considering travel speed, collision avoidance, and survival. Additionally, a dynamic highway environment is configured using seeded randomization of vehicle lanes, initial speeds, longitudinal spacing, and selected behavioral parameters, thereby generating diverse and reproducible traffic conditions. Each method is trained over five independent runs and evaluated deterministically under identical traffic seeds. The results show that incorporating risk assessment improves the success rate of both policies while reducing collisions and unnecessary overtaking. Comparative evaluations demonstrate that proximal policy optimization offers conservative behavior, often resulting in lower-speed navigation, whereas deep Q-network exhibits faster and more aggressive strategies that favor high-speed performance.