Autism spectrum disorder (ASD) is a neurodevelopmental condition characterized by deficits in social communication and restricted repetitive behaviors. Over the past two decades, ASD prevalence has risen substantially. Current estimates suggest that approximately one in 36 children in the United States is diagnosed with ASD [1]. A similar trend has been observed in South Korea, where a population-based study reported a prevalence of 2.64% [1,2]. This rising prevalence underscores the need for early identification and intervention, particularly during the preschool years, when developmental plasticity is high [3].

ASD was traditionally considered a lifelong condition, but longitudinal research has revealed more varied developmental outcomes [4]. While most children with ASD continue to meet diagnostic criteria over time, some show sufficient developmental progress to no longer qualify for the diagnosis; this may be termed ‘desisting’ or ‘loss of autism diagnosis (LAD)’ [4,5]. Conversely, some children appear subclinical early on but later develop clear ASD features, a pattern often described as late-diagnosed ASD [4]. These different trajectories underscore the importance of repeated assessment during early development, especially in young children with borderline or atypical presentations [6]. Several factors have been linked to diagnostic change, including early symptom severity, language and cognitive ability, social responsiveness, and access to early intervention [4,5,7]. Higher early language skills and cognitive functioning generally predict more favorable outcomes [4,5,7-10]. However, most research on these predictors comes from outside Korea. Differences in diagnostic practices, early childhood service structures, and broader cultural context may limit the direct applicability of these findings to South Korean clinical practice [11].

In South Korea, the Childhood Autism Rating Scale (CARS), standardized as the K-CARS, is one of the most widely used tools for assessing ASD in clinical practice [12,13]. Although the Autism Diagnostic Observation Schedule (ADOS) and Autism Diagnostic Interview-Revised (ADI-R) are considered the gold-standard diagnostic instruments, they are less commonly used in routine care because they are time- and resource-intensive [14]. In contrast, the K-CARS is developmentally appropriate and efficient, shows good diagnostic agreement with the ADOS, and is therefore the primary tool for ASD assessment in many South Korean clinics [14]. However, longitudinal studies using the K-CARS are limited, leaving an important evidence gap in understanding how diagnostic status changes over time and which developmental characteristics predict these changes in real-world clinical populations.

We examined diagnostic trajectories among children evaluated with the K-CARS at a tertiary center. We categorized children into persistent, emerging, and desisting ASD groups and analyzed developmental features to identify predictors of diagnostic change. Our goal was to better understand variability in early ASD diagnosis using longitudinal data from routine clinical practice and to help inform more individualized care strategies for these children.

A total of 295 CARS assessments were conducted during the study period. Sixty-six children underwent assessment more than once, and 43 met the inclusion criteria by scoring above the ASD diagnostic cutoff on at least one occasion. Among these 43 participants, 17 were assessed with the K-CARS at baseline and the K-CARS2-ST at follow-up, whereas the remaining 26 underwent the K-CARS2-ST at both time points. No participant was evaluated using the high-functioning version. These 43 participants (32 boys, 74%) were included in the final analysis. The median age at T1 was 2.9 years (IQR, 2.4 to 3.5), and the median age at follow-up (T2) was 4.3 years (IQR, 3.7 to 5.7). All participants completed developmental assessments at both time points, and most received language interventions during the follow-up period (93%). Baseline characteristics are summarized in Table 1.

At T1, 22 children (51%) met the ASD diagnostic criteria, whereas at T2, 36 children (84%) met these criteria (Fig. 1). Based on diagnostic status at T1 and T2, participants were categorized into three groups: 15 (35%) with persistent ASD, 21 (49%) with emerging ASD, and seven (16%) with desisting ASD.

This study demonstrated heterogeneous diagnostic trajectories of ASD in early childhood based on repeated K-CARS assessments. Only 35% of participants maintained a consistent diagnosis (persistent ASD); 49% were diagnosed later (emerging ASD), and 16% did not meet criteria at follow-up (desisting ASD). The desisting group showed a distinct developmental profile, with higher baseline VMI scores and better outcomes in VMI, SQ, and language at follow-up. In contrast, the emerging group initially exhibited slightly higher SQ and VMI scores than the persistent group; however, these differences disappeared over time. The desisting group showed no significant changes across developmental domains. In contrast, the persistent and emerging groups experienced significant declines in SQ, and the emerging group additionally showed a decline in VMI. In exploratory analyses, baseline VMI and SQ showed trend-level associations with later diagnostic trajectories.

The diagnostic distribution observed in our study was similar to that reported in the population-based study by May et al. [4]. May et al. [4] followed 215 children with parent-reported ASD between 6 and 12 years of age: 37% had persistent ASD, 42% had late-diagnosed ASD, and 18% had desisting ASD. This finding suggests that the longitudinal patterns we identified using the K-CARS align closely with those reported in larger community samples. In contrast, our results differed substantially from those of Kleinman et al. [16], who—after excluding children classified as having non-ASD at both time points—reported 68% persistent, 4.5% emerging, and 27% desisting cases using the same tool (CARS). This discrepancy likely stems from differences in study populations rather than the assessment tool. Kleinman et al.’s [16] cohort comprised toddlers identified through ASD-specific early screening using the Modified Checklist for Autism in Toddlers (M-CHAT), which yielded a more diagnostically stable sample. In contrast, both our tertiary-referral cohort and May et al.’s [4] population-based cohort included children with more varied developmental profiles, which may explain the greater diagnostic variability observed.

Despite this variability in trajectories, diagnostic stability among children initially diagnosed with ASD was relatively high: 68% retained the diagnosis at follow-up. This aligns with Diagnostic and Statistical Manual of Mental Disorders (DSM)-based clinical studies reporting 70%–90% stability and with a systematic review of ADOS-based assessments reporting approximately 75% stability [17-20]. These findings suggest that the K-CARS, despite relatively low sensitivity [21], can reliably detect ASD characteristics in settings where resource-intensive instruments are not practical. However, the somewhat lower stability compared with studies that incorporated clinical judgment [17-20] suggests that CARS-based diagnosis alone may not fully capture the autism spectrum. This limitation underscores the need to integrate clinical evaluation and to conduct repeated assessments, especially in children with borderline initial presentations.

A key contributor to the variability described above was the large number of children classified in the emerging group. Although these children did not meet the diagnostic cutoff at baseline, 71.4% had initial CARS scores of 24.5 or higher, placing them in the ‘spectrum range’ [21]. Across the entire cohort, 37 of 43 children had baseline CARS scores of 24.5 or higher, whereas the remaining six—all of whom were later classified in the emerging group—had scores below 24.5 at baseline. This pattern suggests that although most children in the emerging group already exhibited subthreshold features captured within the spectrum range, a subset initially showed symptoms subtle enough to fall below even this broader classification. Subthreshold presentations are common in early childhood, particularly when restricted and repetitive behaviors are subtle or delayed [22]. The emerging group also had significantly higher baseline SQ scores than the persistent group, suggesting that relatively intact early social functioning may have concealed autistic traits and delayed recognition, as noted in earlier studies [22]. Their SQ and VMI scores decreased over time, similar to those of the persistent group. This finding suggests that emerging ASD in our cohort likely reflects delayed recognition rather than a distinct developmental trajectory. Overall, these findings underscore the importance of repeated evaluations, particularly in children with subtle initial profiles.

Although the desisting group no longer had CARS scores above the diagnostic threshold at follow-up, their clinical outcomes should be interpreted with caution. They showed more favorable developmental outcomes than the other groups, although these gains were not statistically significant, suggesting attenuation of autistic features rather than a fundamental shift in developmental trajectory. Nonetheless, their overall pattern of improvement at follow-up suggests a comparatively favorable functional profile within our cohort. These features are consistent with those reported for LAD, observed in 9%–13% of children diagnosed in early childhood [8,23,24], which is typically associated with mild initial symptoms, a marked decline in CARS scores, and broad functional gains over time [8,25]. The desisting group also shared features with pervasive developmental disorder not otherwise specified profiles described in the DSM-IV, which are associated with better developmental outcomes than classic autism; this suggests that at least some individuals may represent milder ASD phenotypes rather than true recovery [26]. Given the relatively short follow-up period, it remains unclear whether these changes reflect sustained symptom reduction or transient diagnostic fluctuations, underscoring the need for long-term monitoring even in children who no longer meet diagnostic thresholds.

In the predictor analysis that included age at the initial assessment, higher baseline VMI displayed a trend toward an association with desisting ASD status (OR, 3.14; P=0.06), whereas higher baseline SQ tended to be associated with emerging ASD (OR, 2.59; P=0.07); however, neither association reached statistical significance (Table 3). Although early social and communication skills are often regarded as favorable prognostic indicators [27-29], higher early SQ in our study may have reduced the apparent severity of ASD symptoms and contributed to delayed diagnosis rather than indicating better outcomes. This may reflect the fact that SQ, as a broad measure of adaptive functioning, is not particularly sensitive to subtle and context-dependent aspects of reciprocal social behavior [30]. In contrast, VMI develops rapidly during early childhood and may more sensitively reflect neurodevelopmental integrity at this age, potentially serving as a proxy for broader neurodevelopmental integrity [26,31,32]. Supporting this interpretation, Sutera et al. [26] reported that early fine motor skills, rather than symptom severity or communication scores, distinguished toddlers who later did not retain their ASD diagnosis. Similarly, in our study, the desisting group had high baseline VMI scores and a substantial reduction in CARS scores over time, suggesting that intact early visual-motor abilities may indicate developmental resilience and predict diagnostic shifts. Although age at initial assessment did not significantly predict diagnostic change in our sample, future studies with larger cohorts may benefit from examining age-specific subgroups.

This study has several limitations. First, the retrospective design and small cohort size limit the generalizability of our findings. Diagnostic grouping relied solely on K-CARS cutoffs, which, despite their clinical utility, may not fully align with gold-standard tools such as the ADOS and ADI-R. The K-CARS and K-CARS2-ST share highly similar structures and scoring frameworks [33], and validated cutoffs specific to each instrument were applied based on Korean standardization studies (28 for the K-CARS and 30 for the K-CARS2-ST) [13,21]. However, minor variability related to the use of different instruments across time points cannot be excluded. A sensitivity analysis using a unified 30-point cutoff, as recommended in the original English versions of both instruments [33], affected only one participant, suggesting minimal impact on the overall findings. The relatively short follow-up period also limits conclusions regarding the long-term stability of diagnostic changes. Furthermore, although the follow-up interval did not differ significantly between groups, the emerging group had the longest median interval. This variability may have influenced developmental outcomes, underscoring the need for additional research with more consistent follow-up schedules. Intervention types did not differ significantly across groups; however, data on intervention intensity and other detailed characteristics were not available, limiting interpretation of their potential influence. Finally, social functioning and communication were assessed only using SQ and verbal language DQ, which may have limited our ability to detect subtle differences in nonverbal or pragmatic communication that could influence developmental trajectories.

In summary, this study demonstrated heterogeneous diagnostic trajectories of ASD in early childhood, with considerable variability across groups but relatively high stability among children initially diagnosed with ASD. Visual-motor abilities were associated with more favorable diagnostic trajectories, whereas relatively preserved early social functioning may have obscured ASD symptoms and contributed to delayed diagnosis. Because longitudinal research using the K-CARS is scarce, these findings help address an important gap by providing real-world data on diagnostic changes in clinical settings. These findings highlight the importance of clinical judgment and repeated evaluations; however, our results should be interpreted cautiously as preliminary and warrant confirmation in larger cohorts with longer follow-up.