---
title: Case Study Template 1 - Biomedical AI

authors:
- Charlotte Merzbacher
- Achille Fraisse
- Ella Davyson
- Thibaut Goldsborough
- Nikitas Angeletos Chrysaitis
- Sebestyén Kamp
- Passara Chanchotisatien
- Scott Pirrie
- Wolf de Wulf
- Evgenii Lobzaev

---
###### tags: `cdt-case-study`, `biomedical-ai`

# Case Study Template 1 - Biomedical AI

## Genetic Disorder Diagnosis from Photos <!-- A title to identify the case study -->

 <!-- Add link to the case study's hero image here -->

## Summary <!-- One or two sentences summarising the case study. -->

The project involves the development of an application and associated algorithms that can diagnose genetic disorders from photos. The app will use various machine learning algorithms to analyze facial features from photos and identify patterns that are associated with different genetic disorders. This case study raises issues of algorithmic bias, genetic counselling, and acceptable false-positive rates.

## Project Description <!-- A more detailed description of the case study, providing relevant contextual information about the domain (e.g. prisons, courts, use case (e.g. predicting sentencing decisions) and stakeholders (e.g. judges). -->
The genetic disorder diagnostic app will run on a smartphone and give feedback to the user based on analysis of supplied facial photographs. The photographs could be of the user or anyone else the user chooses. The app will be generally available so could be used by anyone, not just medical professionals. The feedback will present an assesment of the presence of various genetic conditions in the person portrayed in the photograph. The feedback may include an estimate of associated confidence.

Recent developments in image analysis have enabled improved facial recognition and analysis software. This software could be used to detect genetic disorders such as Down and Turner syndrome from facial features in images taken of a patient. Over 700 genetic disorders including Williams and Noonan syndrome ([Source](https://www.frontiersin.org/articles/10.3389/fgene.2018.00462/full)), [Source](https://science.howstuffworks.com/life/genetic/dysmorphology.htm) are known to be associated with changes to facial features, such as, shape of chin, lip demarcation, retrusive chin etc. Doctors commonly use these facial cues to identify potential candidates for further genetic testing.

The project will use a dataset of facial images of individuals with known genetic disorders, which will be used to train a machine learning model to identify facial characteristics that are indicative of genetic disorders. The model will also be trained on images of individuals who do not have any of the disorders being classified. The model will be optimised to accurately detect these facial features in a user's photo and provide a diagnosis of the potential genetic disorder.

Several ethical issues must be considered in this project. The use of facial recognition technology raises concerns about privacy, especially when the technology is used to identify individuals with genetic disorders. It is crucial to ensure that user data is protected and that their images are not misused or mishandled.

Another ethical issue to consider is the potential for harm. While early detection of genetic disorders can lead to better patient outcomes and quality of life, there is also the possibility of negative psychological effects resulting from a diagnosis.

Additionally, the project team must be mindful of potential biases in the dataset and the machine learning model. If the dataset used to train the model is biased, it may result in inaccurate or unfair diagnoses. The team must ensure that the dataset used for training is representative and inclusive of diverse populations to mitigate this issue.

Finally, it is essential to consider the potential social implications of the application. The use of facial recognition technology for genetic disorder detection may have implications for social and cultural attitudes towards individuals with genetic disorders. It could also be re-purposed for malicious intent.

## Technology Description <!--- A specific description, in plain language, about the relevant technologies and techniques employed in the case study. -->

The project aims to develop a mobile application that can detect genetic disorders such as Down syndrome by utilizing machine learning facial recognition algorithms to identify facial phenotypes. The application will be designed to capture a user's photo and analyze it using an artificial intelligence model trained on a dataset of facial images of individuals with known genetic disorders. The model will be optimized to accurately detect these facial features in a user's photo and provide a diagnosis of the potential genetic disorder.

To evaluate the accuracy of the application, the project will conduct a series of tests using both real and synthetic data. Real data will involve obtaining facial images from individuals with known genetic disorders, while synthetic data will involve generating images of faces with artificially induced genetic disorders. The results of these tests will be used to refine the machine learning model and improve the performance of the application. This could be quantified using precision and recall measures and possibly associated confidence measures.

The ultimate goal of the project is to create an accurate and user-friendly mobile application that can be used by healthcare professionals and individuals alike to detect genetic disorders through facial recognition technology. By providing an accessible and affordable means of screening for genetic disorders, the application has the potential to significantly improve early diagnosis and intervention, leading to improved patient outcomes and quality of life.

Examples of this implementation include:

Hong, D., Zheng, YY., Xin, Y. et al. Genetic syndromes screening by facial recognition technology: VGG-16 screening model construction and evaluation. Orphanet J Rare Dis 16, 344 (2021). https://doi.org/10.1186/s13023-021-01979-y

Gurovich, Y., Hanani, Y., Bar, O., Nadav, G., Fleischer, N., Gelbman, D., Basel-Salmon, L., Krawitz, P. M., Kamphausen, S. B., Zenker, M., Bird, L. M., & Gripp, K. W. (2019). Identifying facial phenotypes of genetic disorders using deep learning. Nature Medicine, 25(1), 60–64. https://doi.org/10.1038/s41591-018-0279-0

Home. (n.d.). Face2Gene. Retrieved March 2, 2023, from https://www.face2gene.com/

## Key Issues <!-- A list of key issues that the case study gives rise to (e.g. data security or privacy concerns). Simple tags rather than sentences (e.g. data privacy concerns). -->

* The discriminatory impact of classification algorithms needs to be investigated.
* Due to the rarity of certain conditions, there may be limited data points available for training the model, particularly across different races and genders.
* The feasibility of incorporating multiple disorders into a single model should be examined.
* False positives generated by the model may cause distress for users and require attention.
* Strategies to prevent the submission of unauthorized photos should be implemented.
* The question of who has access to the model's results and how they will be used must be considered.
* The tight link between facial features and ethnicity may introduce bias into the algorithm and necessitates special attention.
* The issue of accountability and responsibility in cases of misdiagnosis must be addressed.
* Biases can arise in the algorithm if the dataset used for training is unrepresentative.
* The quality of the input images may influence the accuracy of the model's diagnoses.
* Ensuring transparency in the model's decision-making process is essential, including understanding how it arrives at particular conclusions.
* Maintaining anonymity in datasets for facial recognition technology can be a challenge.
* The process of fine-tuning the model requires careful consideration, including who decides on the approach.

<!--
* Discriminatory impact of classification algorithm.
* Many of these conditions are rare so there will be few data points to train model (especially spread across race and genders)
* Too many disorders. Can they be used in a single model.
* False positives may be distressing for users
* How to stop people submitting other peoples photos?
* Who has access to the result?
* Who is going to use the app?
* Facial features are often tightly linked to ethnicity, so algorithms are prone to be biased.
* Accountability / responsibility in the case of a misdiagnosis
* Unrepresentative dataset can lead to bias in the algorithm?
* Image quality has an impact on diagnosis?
* transparency of the model? It's important to know how the model arrives to certain conclusions
* Facial recognition technology is extremely advanced, so making datasets entirely anonymous will be challenging if not impossible.
* Fine tuning of the model, how is it done? Who decides?
-->

## Deliberative Prompts <!-- A series of questions that can be used to help reflect on and deliberate about the key issues. -->

1. In what matter should the data be gathered to guarantee that informed consent is obtained?
2. How will the collection of training data be performed, and how can it be ensured that it is representative of a diverse population in terms of gender, race, ethnicity, etc.?
3. How can facial recognition algorithms be developed in a way that is unbiased, inclusive, and respects individuals' privacy rights?
4. What level of precision is deemed appropriate for the system?
5. Is there a possibility that self-diagnosis using the app could result in increased stress or costs for healthcare systems?
6. Is the diagnostic accuracy of this approach superior to that of genetic screening tests?
7. What is the most effective way to present the results to users, given their understanding of probabilities and the implications of a given percentage likelihood of having a particular condition?
8. What safeguards can be put in place to prevent misuse of the technology or discriminatory practices, such as the identification and stigmatization of individuals based on their genetic makeup?

<!--
1. How should the data be collected to ensure informed consent is obtained?
2. How will the training data be collected? How will it ensure representation across gender, race, ethnicities, etc. so as to be representative of wider population
3. What is an appropriate level of accuracy for the system?
4. Is there a risk that 'self diagnosis' using the app might lead to increase stress/costs for health care systems?
5. Is this type of diagnosis better than genetic screening test?
6. How should the results be presented to the user? Do people understand enough out probabilities for example? What does it even mean if someone is 38% likely to have condition X?
-->

## Stakeholders and Affected People <!-- A list of the project's stakeholders and any people that are affected by the use of the system (e.g. users or communities). -->

* Patients who may be diagnosed with genetic disorders using the mentioned facial recognition technology.
* Users who provide their photos to the system for diagnosis.
* Healthcare professionals who may use the technology in the diagnostic process.
* Individuals who have provided their data for the purpose of the development of the technology.
* Researchers and developers who are involved in creating and refining the facial recognition algorithms used in the technology.
* Regulatory bodies and policymakers who may oversee the ethical use and development of the technology.
* Privacy advocates and civil liberties organizations who may have concerns about the use of facial recognition technology for medical diagnosis and genetic profiling.
* The broader public, who may be affected by the implications of this technology for society, such as the potential for discriminatory practices or the impact on healthcare costs and resource allocation.

<!--
* Patients
* Carers
* Doctors
* System developers
* Health insurers
* Patient' relatives and family in the case of hereditary disorders
-->

## Datasheet

### Available Data <!-- A list of the data types that could be collected, analysed, and used at various stages across the project lifecycle. These could be permissible or impermissible, as determined by the users of the case study. -->

* Positive genetic test results for variety of conditions
* Photos of people's faces with each condition
* Photos of people without any of the genetic conditions that the model is being trained to detect
* Prevalence of each condition (might be needed for certain techniques) - this can be used to calibrate model probabilities if training samples are not representative.

### Algorithmic Techniques <!-- A list of the possible algorithms that could be employed to support analysis or model development. Multiple techniques could be listed if the case study emphasises model comparison and evaluation. -->

* Algorithms to be used for diagnosis would likely consist of deep neural networks such as Convolutional Neural Networks (CNN), Visual Geometry Group (VGG), Residual Networks (ResNet), Inception, Densely Connected Convolutional Neural Networks (DenseNet), and more.
* There are two main options that can be considered when implementing the methodolodgy:
* Implement one large model which can be used to diagnose all conditions.
* Considerations:
* Can combine cross-information
* Model would likely be uninterpretable
* Ability to take other modalities as input
* Use seperate models that are trained for separate conditions.
* Considerations:
* Treats conditions as independent of one another
* Possibly more interpretable
* Possibly trainable with smaller amounts of data

<!--
* Algorithms:
* Likely some sort of deep neural network e.g. Convolutional neural network
* These are often non interpretable
* Either:
* One large model for all conditions
* Considerations:
* Can combine cross-information
* Probably a deep uninterpretable model
* Can take other modalities as input (e.g. ECG)
* Separate models that are trained per condition
* Considerations:
* Treats conditions as independent
* Possibly interpretable
* Possibly trainable with small amounts of data
-->