Real-world optimization problems require some advanced metaheuristic algorithms, which functionally sustain a variety of solutions and technically explore the tracking space to find the global optimal solution or optimizer. One such algorithm is the newly developed COOT algorithm that is used to solve complex optimization problems. However, like other swarm intelligence algorithms, the COOT algorithm also faces the issues of low diversity, slow iteration speed, and stagnation in local optimization. In order to ameliorate these deficiencies, an improved population-initialized COOT algorithm named COBHCOOT is developed by integrating chaos map, opposition-based learning strategy and hunting strategy, which are used to accelerate the global convergence speed and boost the exploration efficiency and solution quality of the algorithm. To validate the dominance of the proposed COBHCOOT, it is compared with the original COOT algorithm and the well-known natural heuristic optimization algorithm on the recognized CEC2017 and CEC2019 benchmark suites, respectively. For the 29 CEC2017 problems, COBHCOOT performed the best in 15 (51.72%, 30-Dim), 14 (48.28%, 50-Dim) and 11 (37.93%, 100-Dim) respectively, and for the 10 CEC2019 benchmark functions, COBHCOOT performed the best in 7 of them. Furthermore, the practicability and potential of COBHCOOT are also highlighted by solving two engineering optimization problems and four truss structure optimization problems. Eventually, to examine the validity and performance of COBHCOOT for medical feature selection, eight medical datasets are used as benchmarks to compare with other superior methods in terms of average accuracy and number of features. Particularly, COBHCOOT is applied to the feature selection of cervical cancer behavior risk dataset. The findings testified that COBHCOOT achieves better accuracy with a minimal number of features compared with the comparison methods.