Hemodialysis as the primary renal replacement therapy for patients afflicted by end-stage renal disease, necessitates monitoring of electrolyte cations in the dialysate to ensure the safety and efficacy of the treatment. Concerned elements include calcium, potassium, and sodium. Currently, prevailing detection methods fall short in providing in situ and real-time detection capabilities, highlighting the need of innovation. Elemental analysis of bulk liquids with laser-induced breakdown spectroscopy (LIBS) presents an interesting potential for dialysate monitoring. Methodological research should further aim to achieve simultaneous multi-element determinations in dialysates with high accuracy and precision, using an arrangement compatible with a flowing dialysate and ready for deployment in a clinical environment. In this work, we developed an experimental arrangement tailored for dialysate analysis, integrating an optimized LIBS setup into a circulating dialysate system, which simulated the clinical operation. Such an arrangement allowed generating a stable and hot plasma within a flowing dialysate. The emission lines from calcium, potassium, and sodium were subsequently captured by a detection system optimized for a simultaneous detection of the three elements with concentrations corresponding to those of a real dialysate. Due to self-absorption of the lines, notably self-reversal of the sodium D-lines, univariate regression, usually efficient for processing LIBS spectra of electrolyte cations in a liquid, was revealed insufficient to provide an analytical performance satisfying the requirements of the clinical practice. Data processing based on multivariate regression was thus developed in this work to substantially improve the model prediction performances. More specifically, back-propagation neural networks (BPNN) coupled with feature selection demonstrated the effectiveness to harness the complexity of the spectrum heavily deformed by selfabsorption and self-reversal of the characteristic lines. The resulted models exhibit a limit of detection (LOD) of 0.302, 0.571, and 4.25 mmol/L, respectively for Ca, K, and Na. The prediction performances of the trained models were assessed using independent test samples, yielding root mean square errors of prediction (RMSEP) of 0.106, 0.137, and 0.646 mmol/L for Ca, K, and Na, respectively. Such performances satisfy the clinical requirements for monitoring of electrolyte cations in a dialysate.