Background Dementia can impair cognitive functions in older adults and further affect language abilities, such as sentence construction. The use of linguistic biomarkers for detecting cognitive decline and early stages of dementia has demonstrated great potential due to its low-cost and non-invasive nature, especially towards a large aging population.Aims In this study, we focused on the syntactic complexity and aimed to identify highly discriminative syntactic features and those set in differentiating Mandarin-speaking older adults with and without cognitive impairment using machine learning models.Methods & procedures We established a new cohort consisting of 52 Chinese older participants, which has not been reported in any previous studies. After applying exclusion criteria, 49 were selected for this study, including 24 labeled as cognitively normal (CN) and 25 as cognitively impaired (CI) based on their scores in the MoCA-B test (Chinese version). Each of the participants completed three connected speech tasks: a picture description task (the Picnic Scene), a story narrative task (Sanmao’s Wanderings), and a story recall task (Cowherd and Weaver Girl). The speech recordings were automatically transcribed into text files and manually checked. We then performed treebank annotation, which was also manually corrected. Subsequently, 17 linguistic unit-based and 10 syntactic structure-based features were extracted from the texts using Python codes. To identify key syntactic features and feature sets, we first examined the discriminative power of these features per se for distinguishing between the CI and CN groups. Then we employed three machine learning algorithms to identify feature combinations performing the best in terms of high accuracy and AUC in cross-validation and minimal number of basic features.Outcomes & results No single feature or feature panels showed high discriminability between groups across all three tasks. However, certain macro-level features with high measurement robustness, such as mean depth of nodes, mean dependency distance, and the ratio of head-final dependencies in specific tasks are noteworthy. Moreover, through machine learning algorithms, we identified several well-performing feature combinations in terms of the metrics of cross-validation mean accuracy and area under curve (AUC). Notably, a feature panel with only two features (mean depth of nodes, and proportions of prepositional dependencies) achieved the best result across classifiers in the story narrative task (ACC = 0.840, AUC = 0.867).Conclusions This study demonstrated the potential of using purely treebank-derived features within a single linguistic domain (syntax) for cost-effective cognitive impairment screening, indicating the value of syntactic biomarkers for cognitive impairment in Mandarin-Speaking Older Adults.