Chronic Neck Pain (CNP) can be associated with biomechanical changes. This paper investigates the changes in patterns of walking kinematics along a curvilinear trajectory and uses a specially designed feature space, coupled with a machine learning framework to conduct a data-driven differential diagnosis, between asymptomatic individuals and those with CNP. For this, 126 kinematic features were collected from seven body segments of 40 participants (20 asymptomatic, 20 individuals with CNP). The features space was processed through a Neighbourhood Component Analysis (NCA) algorithm to systematically select the most significant features which have the maximum discriminative power for conducting the differential diagnosis. The selected features were then processed by a K-Nearest Neighbors (K-NN) classifier to conduct the task. Our results show that, through a systematic selection of feature space, we can significantly increase the classification accuracy. In this regard, a 35% increase is reported after applying the NCA. Thus, we have shown that using only 13 features (of which 61% belong to kinematic features and 39% to statistical features) from five body segments (Head, Trunk, Pelvic, Hip and Knee) we can achieve an accuracy, sensitivity and specificity of 82.50%, 80.95% and 84.21% respectively. This promising result highlights the importance of curvilinear kinematic features through the proposed information processing pipeline for conducting differential diagnosis and could be tested in future studies to predict the likelihood of people developing recurrent neck pain.