Artificial intelligence in the differential diagnosis of progressive ataxias: clinical applications and comparison with conventional diagnostic methods — a systematic review

Abstract

Background: Progressive ataxias are a group of neurological disorders that impair motor coordination, speech, and eye movements, and may have several hereditary or acquired etiologies. The diagnosis of these conditions is challenging, making the use of Artificial Intelligence (AI) tools increasingly relevant to assist in diagnostic processes. Objective: To describe the diagnostic accuracy of artificial intelligence models for progressive ataxias compared with clinical diagnosis. Methods: This study is a comparative systematic review conducted according to the PICOTS strategy and following the PRISMA 2020 guidelines, with protocol registration in PROSPERO. The literature search was performed in PubMed, Cochrane Library, and SciELO databases, including clinical studies published in English or Portuguese over the last 10 years. After screening, seven studies were included out of a total of 399 identified articles according to the eligibility criteria. Results: Seven studies evaluating artificial intelligence models for the differential diagnosis of progressive ataxias were included. The analyzed modalities comprised speech analysis (2 studies), fine motor movements assessed through video (2), gait analysis (1), motor trajectory in virtual reality (1), and graphical image analysis (1). In speech-based studies, performance was mainly assessed using the Area Under the Curve (AUC), which reflects the discriminative ability of the model between diagnostic groups. AUC values ranged from 0.89 to 0.96 when distinguishing ataxia from parkinsonism or controls. One model achieved an AUC of 0.96 and outperformed medical residents in differentiating ataxic from hypokinetic dysarthria, also showing a strong correlation with clinical severity scales (R² = 0.73). In video-based motion analysis, models achieved AUC values between 0.91 and 0.92 in distinguishing ataxia from parkinsonism, with moderate correlation with clinical scales (r = 0.64). The distinction between essential tremor and cerebellar disorders showed an accuracy of 83.6%. Gait analysis demonstrated an accuracy of 90% and an AUC of 0.90, while virtual reality-based motor assessment reached an accuracy of 94.08%, increasing to 98.33% when synthetic data were used. The image-based study presented more moderate performance (accuracy 73%). Overall, five of the seven studies reported AUC ≥ 0.90 or accuracy ≥ 90%, indicating high diagnostic performance compared with the clinical reference standard. Conclusion: Artificial intelligence models demonstrate consistent diagnostic performance in the differential diagnosis of progressive ataxias compared with the clinical reference standard, particularly in approaches based on dynamic digital biomarkers. Although external validation and methodological standardization are still required, these findings suggest that AI-based tools have relevant potential as complementary support for neurological diagnosis and may also contribute to disease severity monitoring.