Extended Data Fig. 4: Encoding and decoding of medical information on MNP.

(a) During the encoding phase, information data is converted to a pattern that can be encoded on a MNP. Encoding phase compensates for the loss of individual bits of a microneedle patch over time and ensures error correction up to a certain percentage of bit corruption. Once the type of information data to be recorded is determined, it is translated to a binary string and then to information bits. Since the system is prone to unforeseeable errors such as missing bits from environmental trauma, temporal decay of fluorescent dye, and false positive signal from background noise, redundancy is added to the information bits using Reed-Muller error correcting code. Then, the generated string is mapped to a 2D pattern that fits in a template with four corners reserved for orientation. Encoding bits are arranged sequentially from top left to bottom right to generate an initial encoded pattern. An encryption mask is also added to ensure the privacy of personal medical data. (b) Decoding phase correctly translates acquired raw image back to the medical information that was originally recorded on patients during encoding phase. Decoding phase takes potential spatial imperfectness into consideration and makes a robust image recognition system based on deep learning (DL). Raw image is binarized using a DL-based image binarization network. Raw RGB image is converted to a BnW binary image, rectified to an axis-aligned and upright square geometry, and cropped and rotated to identify the MNP region by finding a minimum area rectangle. It is then fed into a DL-based image recognition network. The recognized binary array is re-oriented, and encryption step is reversed. Array is remapped into binary units for the Reed-Muller error correction decoding step and converted back to the corresponding string. Finally, it is translated back to the corresponding medical information text and is retrieved on a screen. (c) RM ECC adds redundancy to information bits so the transmitted message can be accurately recovered even when some bits are erroneously flipped. RM ECC corrects independent, non-block-based binary bits and is a good option for the OPMR system because spatial correlation between individual microneedle bits cannot be assumed for OPMR MNPs, and this would ensure a reliable long-term data retrieval. (d) Adding a known and fixed encryption pattern ensures the privacy of personal medical data of the OPMR system. i. The number of orientation bits are determined by subtracting encoding bits as per RM code from the total number of bits on an MNP. The orientation bits are allocated at four corners of the MNP with the bottom right corner OFF. ii. A pattern is first generated as a 2D array with roughly half ON-bits and half OFF-bits. iii. An encryption mask is added to the initially generated pattern. iv. After randomly flipping pixels on the raw encoded pattern, the encrypted pattern will consist of half ON and OFF pixels on average, which makes the recognition system robust to any patterns during the decoding step.