Robust Machine Learning for Medical Diagnosis: Feature Extraction and Threshold Calibration with Application to Preterm Birth Prediction

Preterm birth (PTB) remains a major challenge in global obstetrics. Early prediction is difficult due to limited data sources, small sample sizes, and class imbalance. Beyond these modelling difficulties, machine learning models frequently suffer from decision breakdown in new environments, as distributional shifts alter score distributions and distort the originally calibrated decision threshold, thereby undermining diagnostic reliability.
This thesis first develops a pipeline for PTB risk prediction using cervical Electrical Impedance Spectroscopy (EIS) as a non-invasive diagnostic modality. The framework integrates feature selection strategies (filter, embedded, and wrapper methods) with data-balancing techniques to enhance discriminative performance. Results demonstrate that EIS captures meaningful biophysical signatures.
Building on this modelling foundation, the thesis addresses a critical yet underexplored challenge: how fixed decision thresholds fail under distribution shift. To bridge this gap, a Threshold-Calibration Decision Process (TCDP) framework is proposed. By exploiting the invariance of ROC under strictly monotonic transformations, TCDP enables decision thresholds to be recalibrated using negative-only target data, without requiring model retraining.
The proposed framework is validated in both EIS-based and medical record based PTB predictions. Experimental results show that the proposed TCDP method improves performance by aligning sensitivity–specificity trade-offs with training optimal point. On the PTB datasets, TCDP consistently achieved the smallest distance metric, reducing it from 0.202 to 0.151 for the MLP model and medical records from 0.372 to 0.180 for the RF model. TCDP achieved a maximum sensitivity of 0.915 and a maximum specificity of 0.960, while reducing the distance to the optimal ROC operating point by 21.4%–90.6% compared with LPM.
This work presents a framework that connects biophysical signal modelling, machine learning methods, and decision threshold calibration. The proposed decision threshold calibration approach can potentially advance the safe and scalable deployment of machine learning systems in different medical diagnosis scenarios.