Conclusion
In conclusion, the project’s aim of designing and constructing a Real-Time Wearable System for Early Detection of Diabetic Foot Ulcer has been very close to being fulfilled. The design phase was completed, the data communication via BLE, data logging and the Machine Learning architecture, i.e., the software section was completed. Unfortunately, there were some unfinished tasks in the hardware section, and the necessary data was not able to be gathered due to unforeseen conditions. The BLE interface was tested many times to check and ensure the robustness of the transmission and reception of the data string. The data logger was tested and verified if it is saving the data intact (without cutting any part of the data string).
Half of the hardware part, which was finished by the group (one PCB soldered and one insole), was tested so that all the sensors and other components are working well.
The Machine Learner’s accuracy, precision and other parameters were not something to boast about, but it provided decent results, especially for the FKNN algorithm. ML’s results for the Diabetic patients were mostly poor due to ambiguity present in this class. If more labeled data could be gathered to train the algorithm, the performance of the ML would be much higher, and high accuracy and precision in detecting or predicting the patient’s class are one of the conditions behind making the system robust.
The temperature sensors were calibrated based on the data provided by the verified manufacturer form LittleFuse, but nevertheless, a calibration setup could be set but was not possible due to the complexity revolving calibrating flexible sensors and lack of time. But the temperature sensor readings were not incorporated in detecting the patient’s class in this case since its implementation should be during detecting real patients, and that data was not available. Nevertheless, it has to be remembered that no ML is needed for the temperature sensor. Combining the result of vGRF from the ML and the temperature sensors will make the outcome more robust, and clinical decisions could be made based on this. The device being so low cost and easy to handle (price might decrease even more if the product goes for mass production after advanced technical and economical testing), it could be used personally. Moreover, the device could help economically insufficient patients in rural areas and in poorer countries, after making it ethically acceptable.
Future Work
The first and foremost task for the team would be to finish the unfinished business about the hardware section of the project as soon as the blockade ends. Then the team can go for data gathering in various hospitals and diabetic centers around Qatar and even abroad (if not enough data could be collected). The team’s primary coordinator, Dr. Muhammad Chowdhury, already received permission from the Qatar Diabetes Association (QDA) to conduct the data collection from various hospitals in Qatar. After gathering enough data (at least 500 subjects from three classes mentioned before), the ML can be trained to get better training accuracy. Then the testing accuracy can be calculated by taking labeled data from the known groups. If the ML can provide good training and testing accuracy from this supervised training, the team can go for predicting patients with unknown conditions and check for the outcome. Now, since there is a stable ML classifier developed already, it can be run for the IMU readings and check for the outcome. After gathering the target patients’ IMU data along with vGRF, if the system provided a good response for at least some of the IMU parameters, this can be considered along with the vGRF outcome for an even more trusted result. But it requires more research and dedication. Finally, a Graphical User Interface (GUI) can be developed to show the result in an interactive manner.