This paper presents the analysis and a proof-of-concept design of the low-complexity early termination hybrid automatic repeat request (ET-HARQ) for ultra-reliable low-latency communication (URLLC). In ET-HARQ, unfit packets are flagged for retransmission prior to decoding, a process that significantly diminishes decoding complexity and facilitates swift HARQ reporting. This characteristic makes ET-HARQ well-suited for URLLC applications. We analyze the impact of ET-HARQ on the packet error rate (PER), throughput, and complexity performance in the finite block length regime, taking into consideration cyclic redundancy check (CRC) limitations. Numerical results indicate that ET-HARQ significantly reduces the decoding complexity and improves throughput with little to no loss in the PER. In addition, ET-HARQ demonstrates resilience even with a short CRC, whereas the imperfections of a short CRC significantly impact PER reliability in standard HARQ. To validate our analysis, we also design a practical early termination mechanism involving belief propagation and neural network (BP-NN) to predict the decodability of the received packet. Testing with BCH and CRC-polar codes shows that it can reach up to a 70 %similar to 80 % prediction accuracy with packet lengths less than 128 bits encoded with high-density linear block codes. Simulation shows that the BP-NN-based ET-HARQ has significantly lower complexity at this accuracy level than the standard HARQ with similar reliability.