*2.2. Dataset*

This section explains the proposed approach to predicting the DNA amount per bead using experimental data and leveraging deep learning methods. Figure 2 shows an overview of the proposed method. In this study, the dataset was obtained from a custom-based microfluidic chip to detect DNA molecules bound to beads by measuring the impedance peak response (IPR) at multiple frequencies [5]. The proposed machine learning method will be trained on the electrical signals obtained from the biosensor with a specific configuration of the channel and electrode size. It is anticipated that increasing the size of the channel decreases the sensitivity of the biosensor. This means that the passage of beads or particles through the electrodes will result in weaker signals, i.e., smaller peaks in the impedance signals. This makes it more difficult to distinguish the passage of beads with very small amounts of DNA from the signal noise. Consequently, the accuracy of the machine learning method will be negatively affected. In the case of using another configuration, it would be more accurate to retrain the model based on the data obtained from the sensor with the new configuration.

**Figure 2.** Overview of the proposed framework for prediction of DNA amount per bead.

The impedance response was measured simultaneously at 8 different frequencies ranging from 100 kHz to 20 MHz [5]. This dataset contains 105,104 data points collected on different days. For each piece of data, the frequency; the real, imaginary, and absolute values of the peak intensity; and the phase change of the peak intensity were measured to calculate the DNA amount per bead. All these features are used as input for the neural network model. In this work, our goal is to find a relationship between the aforementioned measurement features and the DNA amount per bead. To accomplish this, we explored three different machine learning approaches: classification, regression, and a hybrid model. The hybrid model is a combination of the best architecture of the classification and regression models.

The proposed model consists of three main steps, which are shown in Figure 2. The frequency; real, imaginary, and absolute values of the peak intensity; and the phase change of the peak intensity were recorded in measurements and will be used as input features. The output is the DNA amount per bead. In total, 7 outputs were examined containing 6 different concentrations of DNA from low to high coupled to paramagnetic beads and one control bead, which is a bare bead (i.e., a bead with no DNA concentration).
