Biosensor performance is strongly influenced by coupled fabrication variables and measurement conditions, yet optimization often relies on empirical parameter tuning. Here, we present a machine learning-based analytical framework to quantify the relationship between fabrication conditions and electrical responses in an MXene-based extended-gate field-effect transistor aptasensor for human parainfluenza protein detection. A dataset of 35 condition-level sensor responses was constructed from independently fabricated sensors by varying MXene-related, aptamer-related, target-related, and buffer conditions. Six regression models−partial least squares, sparse PLS, elastic net (EN), algebraic learning via elastic net, random forest, and support vector regression were evaluated using predictive accuracy, train-test robustness, and interpretability. Most models achieved test mean absolute percentage error values of approximately 3 %, indicating that the measured electrical responses could be predicted from multivariable experimental conditions. The EN model provided the most balanced performance, combining accurate prediction with robust generalization and interpretable coefficients. Variable-importance and sensitivity analyses identified nanomaterial concentration and target concentration as dominant contributors to the threshold-voltage response. These results demonstrate that machine learning can support systematic interpretation of fabrication-dependent biosensor responses and provide a practical route for data-driven optimization of electrical aptasensing platforms.