**8. Limitations**

Although flexible devices have demonstrated tremendous potential, they are rarely commercialized or used clinically. Low throughput or expensive manufacturing methods make many of these devices difficult to implement widely. For example, non-traditional substrates like PET or TPU cannot be manufactured with large-scale conventional cleanroom fabrication [96,97]. High-throughput manufacturing processes, such as screen printing, have been demonstrated, yet optimal parameters have not yet been discovered. In addition, the resolution of these processes remains low. The elimination of motion artifacts also presents significant challenges. For example, dry electrodes still suffer high impedance with the skin. In addition, powering these devices for long-term monitoring remains a challenge, especially for battery-less devices, and low power consumption is an essential trait for continuous monitoring devices. The addition of wireless transmission also increases the power consumption needed. Bluetooth, for example, consumes up to 5 mW of power, which is more than many thin-film batteries can provide. New machine learning methods can help detect the presence of arrhythmia in heart rates. However, distinguishing between types of arrhythmias and between other classes of heart disease has proven difficult. Machine learning is also limited by the quantity and quality of the training data. In addition, many low-training-data machine learning models are prone to overfitting data and are therefore unable to generalize the testing data. Machine learning is also computationally intensive, making real-time data classification difficult. Finally, the "black-box" nature of machine learning is inherently complex for doctors and clinicians to interpret.
