Toward point-of-care assessment of patient response: a portable tool for rapidly assessing cancer drug efficacy using multifrequency impedance cytometry and supervised machine learning

Abstract

We present a novel method to rapidly assess drug efficacy in targeted cancer therapy, where antineoplastic agents are conjugated to antibodies targeting surface markers on tumor cells. We have fabricated and characterized a device capable of rapidly assessing tumor cell sensitivity to drugs using multifrequency impedance spectroscopy in combination with supervised machine learning for enhanced classification accuracy. Currently commercially available devices for the automated analysis of cell viability are based on staining, which fundamentally limits the subsequent characterization of these cells as well as downstream molecular analysis. Our approach requires as little as 20 μL of volume and avoids staining allowing for further downstream molecular analysis. To the best of our knowledge, this manuscript presents the first comprehensive attempt to using high-dimensional data and supervised machine learning, particularly phase change spectra obtained from multi-frequency impedance cytometry as features for the support vector machine classifier, to assess viability of cells without staining or labelling.

Keywords: Electrical and electronic engineering; Sensors.

Conflict of interest statement

Conflict of interestThe authors declare that they have no conflict of interest.

© The Author(s) 2019.

Figures

Fig. 1. Schematic diagram of the system.

Multifrequency impedance cytometry measures the response across a broad range of frequencies for assessment of cellular response to target drug. Live cells and dead cells are assessed using machine learning algorithm to predict their viability

Fig. 2. Device micrograph.

a Microfabricated electrodes at the channel. b Cancer cells flowing through the microfluidic channel

Fig. 3. Device circuit model.

Equivalent circuit model of the electrode–electrolyte interface in the microchannel along with the readout circuit for measuring changes in resistance across the channel

Fig. 4. Time series data of single cells.

Normalized impedance response of a live cancer cells and b dead cancer cells at 500 kHz, 20 MHz, and 30 MHz. Each peak corresponds to a single cell being detected

Fig. 5. Frequency dependent amplitude change.

Amplitude spectrum of live cancer cells and dead cancer cells

Fig. 6. Multi-frequency analysis for amplitude change.

a Scatter plot of amplitude change for live cancer cells and dead cancer cells at 500 kHz and 20 MHz. b Scatter plot of amplitude change for live cancer cells and dead cancer cells at 500 kHz and 30 MHz

Fig. 7. Classifier accuracy when amplitude change used as features.

a Confusion matrix of the SVM classifier while using amplitude change as features. b Comparison between the analysis of cell viability by Trypan Blue staining method (ground truth) and multifrequency impedance spectroscopy with SVM using impedance change as features for the SVM classifier

Fig. 8. Multi-frequency phase analysis.

a Scatter plot of phase change for live cancer cells and dead cancer cells at 500 kHz and 20 MHz. b Scatter plot of phase change for live cancer cells and dead cancer cells at 500 kHz and 30 MHz

Fig. 9. Classification accuracy when phase change used as features.

a Confusion matrix of the SVM classifier while using phase change as features. b Comparison between the analysis of cell viability by Trypan Blue staining method (ground truth) and multifrequency impedance spectroscopy with SVM using phase change as features for the SVM classifier

Fig. 10. Classification accuracy when both amplitude and phase change used as features.

a Confusion matrix of the SVM classifier while using amplitude change and phase change as features. b Comparison between the analysis of cell viability by Trypan Blue staining method (ground truth) and multifrequency impedance spectroscopy with SVM using amplitude change and phase change as features for the SVM classifier