A Novel Multi-model Data-Driven Ensemble Technique for the Prediction of Retention Factor in HPLC Method Development

• Published in: Chromatographia Vol. 83; no. 8; pp. 933 - 945
• Main Authors: Usman, A. G., Işik, Selin, Abba, S. I.
• Format: Journal Article
• Language: English
• Published: Berlin/Heidelberg Springer Berlin Heidelberg 01.08.2020; Springer Nature B.V
• Subjects: Accuracy; Analytical Chemistry; Artificial intelligence; Calibration; Chemistry; Chemistry and Materials Science; Chromatography; Computer simulation; Correlation coefficients; High performance liquid chromatography; Laboratory Medicine; Model accuracy; Multilayers; Neural networks; Original; Pharmacy; Proteomics; Support vector machines; Verification; Retention factor; Ensemble techniques; Amiloride; Artificial intelligence; Methyclothiazide; HPLC
• ISSN: 0009-5893; 1612-1112
• DOI: 10.1007/s10337-020-03912-0

Reliable simulation of retention factor ( k ) is crucial in high-performance liquid chromatography (HPLC) method development. In this research, three different Artificial intelligence (AI) based models, namely the multi-layer perceptron (MLP), Support vector machine (SVM) and Hammerstein–Weiner (HW) models, were employed as well as three ensemble techniques, i.e., neural network ensemble (NNE), weighted average ensemble (WAE) and simple average ensemble (SAE) to predict k for HPLC method development. In this context, the pH and composition of the mobile phase (methanol) are used as the input variables with the corresponding Methyclothiazide (M) and Amiloride (A) as antihypertensive target analytes. The performance efficiency of the models was evaluated using mean square error (MSE), determination coefficient ( R 2 ), and correlation coefficient ( R ). The results obtained from the single models showed that MLP outperformed the other two models and increased the prediction accuracy up to 1% and 3% for the HW and SVM models, respectively, for the prediction of M. However, for the prediction of A, SVM outperformed the other two models and increased the prediction accuracy up to 7% and 6% for HW and MLP, respectively. In the ensemble technique, the results obtained for the prediction of both M and A demonstrated that NNE increased the performance accuracy by 14% of the single models. Also, NNE proved to be superior to the two linear ensembles and improved the prediction accuracy up to 14% and 2% for SAE and WAE, respectively, for the simulation of M with R 2  = 0.9962 and 0.9949 for both calibration and verification, and up to 9% and 6% for A with R 2  = 0.9606 and 0.9569 for both calibration and verification phases respectively. The overall results depicted the reliability and robustness of both the AI-based models and justified the enhancement capability for ensemble techniques for both the two analytes.