Autism Spectrum Disorder (ASD) is a complex neurodevelopmental condition that affects many people across the globe. To effectively intervene, early detection is essential. One potential solution to address challenges and expedite autism referrals and diagnoses is to leverage Artificial Intelligence (AI) algorithms and models. However, AI systems often face algorithmic bias due to class imbalance. Conventional resampling methods, such as random under-sampling, oversampling, and synthetic data generation, are commonly used to mitigate this imbalance, but often distort the data distribution or introduce noise, leading to overfitting. To address these challenges, this study introduces Rule-Based Screening Sampling (RSS), a patterned, non-synthetic strategy that mitigates class imbalance by selecting high-confidence samples from underrepresented classes, guided by rules and thresholds applied to existing data. Unlike conventional techniques, RSS preserves data integrity while improving the representativeness of the training set. Applied to the Autism AI dataset (the largest behavioural dataset with over 12,000 caregiver-reported cases), RSS effectively mitigated class imbalance, where underrepresented autistic or non-autistic cases led to biased model outputs. Comparative evaluation with conventional sampling methods demonstrates that RSS significantly improves key performance, such as unweighted average recall and accuracy, which are often compromised in imbalanced clinical datasets. While evaluated in the context of autism, RSS represents a generalisable concept that could be adapted to other domains by leveraging domain-specific expert rules or intrinsic data patterns without relying on synthetic data.