Smart data driven predictive model application for wound healing tracking

• Main Author: Neto, Pedro Daniel Ribeiro
• Publication Date: 2021
• Format: Master thesis
• Language: eng
• Source: Repositórios Científicos de Acesso Aberto de Portugal (RCAAP)
• Download full: http://hdl.handle.net/10400.14/34902
• topic: Wound healing; Chronic wounds; Machine learning; Wavelet transform; Cicatrização de feridas; Feridas crónicas; Transformada de wavelet

Chronic wounds affect millions of people around the world. Just in the United States, it is estimated that 6.5 million people suffered from chronic wounds, while in Europe the number is estimated to be around 1.5 to 2 million. The number of chronic wounds in Portugal isn’t well known but at least 14,000 people suffer from leg ulcerations at any given time. Chronic wounds tend to affect people of older age or that suffer from chronic diseases such as diabetes. The increase of the average age of the populations in developed countries, coupled with the increase of diabetes cases will only exacerbate the problem of chronic wound prevalence. Despite all the wide array of innovative and potential treatments, current dressings do not provide any feedback information regarding the wound healing process. Developments in biosensors for wounds are being made, however, the processing of the gathered information is still lacking. Using the computational power of today’s processors, a wound healing application was developed that was able to predict wound healing states, infected vs. non-infected, using only inexpensive sensors, thermal images, simple signal processing techniques, and features. Data was collected from 3D skin models and processed using Wavelet transform a powerful tool used in signal analysis allowing the decomposition of humidity and temperature signals in its frequency, even at the low sampling frequency. Features were collected from both humidity, temperature signal, and thermal images, and were selected through a process of feature selection and then feed to machine learning algorithms. It reached a maximum accuracy of 85.7% using a combination of temperature and humidity features feed to a logistic regression algorithm, as well as a Convolutional Neural Network, demonstrating the viability of this method.