Mild Cognitive Impairment (MCI) is a transitional stage between normal aging and dementia, where early detection is crucial. Neuropsychological test batteries, including the Consortium to Establish a Registry for Alzheimer's Disease (CERAD), are widely used but require lengthy administration and involve trained professionals. Digital screening tools with serious game elements may provide an efficient alternative by simulating daily-life scenarios. This study evaluated the diagnostic performance of a digital serious game and proposed strategies for its optimization. Seventy-eight older adults (58 healthy controls, 20 MCI) completed the CERAD-K2 and the digital tool. The serious game consisted of five tasks reflecting daily life: recalling sale items, solving puzzles, crossing malfunctioning traffic lights, a prize (go/no-go) game, and route-finding. Diagnostic performance was evaluated using Random Forest classification (accuracy, sensitivity, specificity), and discrimination was assessed with Receiver Operating Characteristic (ROC) analysis with Area Under the Curve (AUC). Task contribution was examined by removing components, and usability was measured using the NASA Task Load Index (NASA-TLX). The digital tool showed an accuracy of 82.1%, sensitivity of 71.4%, and specificity of 84.4%. Compared with CERAD-K2 (sensitivity 87.9%, specificity 66.7%), the digital tool demonstrated higher specificity but lower sensitivity. Removing the puzzle task improved sensitivity to 76.9% while maintaining specificity at 84.6%. ROC analysis indicated excellent discrimination (AUC=1.0). NASA-TLX results indicated a moderate workload, supporting feasibility. Compared to conventional neuropsychological batteries, the digital tool demonstrated higher specificity but relatively lower sensitivity, limiting its use as a screener. Future optimization should focus on removing tasks that negatively affect sensitivity and emphasizing those with higher feature importance. Specifically, we plan to (i) analyze misclassification patterns by comparing false negatives and true positives, (ii) remove low-importance features with retraining to assess performance changes, and (iii) refine UI/UX to reduce workload and improve clarity, followed by validation in an expanded sample.