STICKER

Inspiration

Countless patients with potential sexually transmitted infections, who should visit a doctor, do not because of a negative stigma. On top of that, numerous patients cannot afford to go to a doctor because of financial or time constraints. Many of these patients, then, become severely ill to the point where they are forced to visit an emergency room.

These patients would greatly benefit if they had an easy access to a consultation which would give them preliminary awareness about their condition and urge them to visit a doctor’s office, if and as necessary.

Even for patients who can visit a doctor’s office, without worrying about financial or time constraints, can see benefit in educating themselves before going to a doctor.

With these factors in mind, we feel that there is a great need for a platform where patients can have an access to not only a consultation but also to the necessary self-education for STIs.

We have, therefore, developed a prototype website, named STIcker, where patients can have some preliminary consultation and self-education. STI in this name stands for “Sexually Transmitted Infection”, C for “classification”, K for “knowledge”, E for “Eradication”, and R for “Reporting”.

Image Classification technology has now entered many sectors such as ATMs and retail stores. It has a great potential for categorization of dermatological diseases, including STIs.

What it does

The “Classification” page of the STIcker website relies on deep learning. If a person uploads several pictures of skin conditions, the website classifies the images and gives a preliminary result, picture by picture, and on average, indicating whether the images imply chlamydia, gonorrhea, syphilis, trichomoniasis, or none. Since we did not have access to an adequate number of photographs, our trained predictive model is not very accurate. But, if given thousands of photographs of each type of STI, it has a potential to become a powerful tool for doctors to use to classify STIs.

Skin images are classified based on skin tones, uniformity, and coloration. For detecting skin problems, the image classifier model should place emphasis on uniformity and coloration, and ignore skin tones.

In training a model from images of items like chairs and tables, the images are augmented by rotating or flipping. In training a model from images of skins, the images can be augmented by changing the skin tone, while maintaining the same skin condition. This method of augmentation would help reduce the bias projected on various skin tones.

Additionally, in order for the deep learning model to be successful in predicting between healthy and unhealthy skin, one must take account of all sorts of skin diseases as well as healthy skin so as to accurately classify STIs. Thus, one must feed the deep learning model many photographs of various skin diseases. And so, this part of the project can be expanded upon to fully account for all types of dermatological conditions.

In a future version, we can allow the image classification anonymously, without uploading data and images to the server. This can help reduce concerns related to HIPAA violations.

The “Knowledge” page of our website provides preliminary information and links about STIs that can help bridge health barrier gaps. Currently, we only have documented resources and information. We have provided a link to a chatbot, where visitors can address specific questions that one may have about STIs. The chatbot can play a significant role here in fostering a more comfortable, non-judgemental discourse for a topic associated with much stigma.

The “Eradication” page follows a model similar to plannedparenthood.org and findhelp.org, wherein by entering one’s interested service and zip code, they are able to find a list of nearby providers that can perform various services including STI testing, diagnosis, and treatment. We hope that through this page, various people of different socioeconomic classes, including the underserved communities, can find and gain access to these health resources, thereby generating greater health equity.

On the “Reporting” page, we want to create more awareness in terms of tracking the number of STIs in a particular location. Though the page has not yet developed filters for accurate reporting, it is possible to incorporate filters at the data intake point. It should be noted that the audience for this page is planners. This allows governmental health programs to be better targeted to areas that have heightened occurrences of STIs. As you can see, an important collaboration between patients, providers, and planners is a key to the success of healthcare.

How we built it

For the Classification page, we gathered some skin images from he Internet, and fed them into a Python and TensorFlow based deep learning image classification program. This gave us an image classification predictive model.

For the Education page, we gathered and posted some information about STIs, and used BotPress for a chatbot.

Challenges we ran into

The first and most prominent challenge we faced was the collection of images for the classification model. Because of the lack of accurate photos available on the web, we were not able to feed the Deep Learning AI model the thousands of photos it requires to accurately predict skin lesions. As one can imagine, this led to an inaccurate prediction model.

Another challenge we ran into was during the process of creating an Android app. Due to variable factors, especially time constraint, an Android app was not finished.

Accomplishments that we're proud of

We are proud that we got a good starting AI, predicting model that can be further trained down the road to become a more accurate classifier.

What we learned

Lalit learned about STIs. Sean learned about AI and Deep Learning model.

What's next for STICKER

There are four additional features we would like to add to STICKER.

1) We would like to collect more photos so as to train the prediction model to become more accurate. 2) We would like to use advanced image augmentation methods to remove skin tone biases. 3) We would like to develop Android and iOS app to enhance the image classification feature. 4) We would like to anonymize the data of photographs so as to avoid HIPAA violations.