To construct a saliva-based caries risk assessment model, saliva samples from 176 severe early childhood caries (S-ECC) children and 178 healthy (H) children were screened by real-time PCR-based quantification of the selected species, including Streptococcus mutans, Prevotella pallens, Prevotella denticola and Lactobacillus fermentum. Host factors including caries status, dmft indices, age, gender, and geographic origin were assessed in their influence on abundance of the targeted species, which revealed host caries status as the dominant factor, followed by dmft indices (both P < 0.01). Moreover, levels of S. mutans and P. denticola in the S-ECC group were significantly higher than those in the healthy group (P < 0.001 for S. mutans and P < 0.01 for P. denticola). Interestingly, the co-occurrence network of these targeted species in the S-ECC group differed from that from the healthy group. Finally, based on the combined change pattern of S. mutans and P. pallens, we constructed an S-ECC diagnosis model with an accuracy of 72%. This saliva-based caries diagnosis model is of potential value for circumstances where sampling dental plague is difficult.