RAPID DIAGNOSIS OF PERITONEAL DIALYSIS-RELATED PERITONITIS BY POINT-OF-CARE VIS/NIR OPTICAL SPECTROSCOPY AND MACHINE LEARNING ALGORITHMS

Peritoneal dialysis (PD)-related peritonitis remains a significant complication among end-stage kidney disease (ESKD) patients, contributing to morbidity, technique failure, and mortality. Current diagnostic workflows rely on peritoneal fluid white blood cell (WBC) counts and culture-based identification, which take 18–72 hours. This delay impedes timely antibiotic administration, with each hour's delay increasing failure risk by 6.8%. A rapid, point-of-care diagnostic modality is urgently needed, particularly for home-based PD management.

We developed a portable diagnostic platform utilising visible/near-infrared (Vis/NIR) transmission spectroscopy, integrated with machine learning algorithms, for the real-time, reagent-free detection of peritonitis. An 18-channel Vis/NIR spectrometer was used to acquire 257 spectra from 28 PD effluent samples collected from 15 patients in real-world conditions. Transmission was measured directly using standard 120 mL polypropylene containers. Spectral preprocessing included baseline correction, standard normal variate normalisation, and smoothing.
Machine learning models were built using Partial Least Squares Regression (PLS) for feature extraction and logistic regression for binary classification (WBC >100 cells/µL). Multi-class classifiers were also implemented to capture patient-specific spectral profiles. Models were trained using stratified, grouped cross-validation and optimised via grid search for macro F1-score. Standard operating procedures (SOPs) for clinical integration and user handling were developed and refined through feedback from dialysis nurses.

The anomaly detection model achieved a sensitivity of 0.933 (95% CI 0.876–0.969), a specificity of 0.854 (95% CI 0.802–0.897), and an F1-score of 0.892, outperforming conventional urine strip and leukocyte esterase tests. Latent space analysis demonstrated strong intra-specimen spectral clustering and inter-patient separability, indicating robust signal quality and generalisability. Dialysate spectra exhibited higher transmittance and intrinsic fluorescence, especially in UV-excited bands, differentiating them from urine profiles. Handling variability from jar repositioning introduced minimal spectral noise, confirming system robustness. The comparative analysis further validated device stability (<1% CV in fixed-jar urine scans) and biofluid-specific spectral heterogeneity patterns.
Pilot clinical feedback confirmed the workflow’s feasibility, with minimal training required for nursing staff and rapid sample-to-result turnaround. 

This study demonstrates that Vis/NIR transmission spectroscopy, combined with machine learning, provides a feasible, rapid, and non-invasive approach for diagnosing PD-related peritonitis. The portable platform shows strong diagnostic performance and operational robustness in real-world clinical settings, laying the foundation for point-of-care implementation. Future directions include cohort expansion and regulatory scale-up for deployment across inpatient and community care settings. This innovation aligns with national efforts toward decentralised healthcare and may reduce hospitalisation, technique failure, and healthcare costs associated with PD-related infections.