Early detection of autism spectrum condition is crucial for children to maximally benefit from early intervention. The study examined a machine learning model predicting the increased likelihood for autism from wellness records from 0 to 24 months. The study included 591,989 non-autistic and 12,846 autistic children. A gradient boosting model with a threefold cross-validation and SHAPley additive explanation tool quantified feature importance. The model had an average area under the curve of 0.81 (SD = 0.004). The high-likelihood group detected by the model had a 0.073 autism spectrum condition incidence rate; 3.42-fold more than in the entire cohort (0.02). Sex-specific models had higher specificity (0.81 boys and 0.79 girls) than sensitivity (0.64 boys and 0.66 girls). The common predictors were more parental concerns, older mothers, never nursing, lower initial and higher last weight percentiles, and several delayed milestones. SHAPley additive explanation tool results show common, important predictors in the full sample and separate boys' and girls' models. These included birth, growth, familial, postnatal parameters and delayed language, fine motor, and social milestones from 12 to 24 months. Machine learning algorithms can help detect increased autism signs by relying on the multidimensional data routinely recorded during the first 2 years.