Bias In Orthodontic AI: Recognizing And Mitigating Disparities

Diese 7 Schritte machen Deine Praxis lean

Every patient is entitled to the best personalized medical care, regardless of caste, color, and personal traits. However, bias exists in all healthcare setups due to differences of opinion. Human decision-making (by healthcare providers) potentially leads to bias and discrimination. It was anticipated that the introduction of artificial intelligence would eliminate bias. But has AI been capable of sparing the medical/dental field from bias? Let’s have a look at how AI identifies and mitigates disparities in treatment.

Cause Of Bias In AI Healthcare

AI tools are extremely efficient and useful. But despite all the perks AI algorithms are prone to bias. This is because AI applications
are the result of deep learning from provided studies. AI implementation in healthcare fields like orthodontics comes from the analysis of multiple studies.

As a lot of studies (conducted by humans) have some aspect of bias, the AI system learning from them suffers the same fate. For example, there has been an underrepresentation of women in clinical trials for heart disease while cardiovascular disease remains a major cause of female mortality. This leads to a lack of cardiac care for women by the healthcare providers. Based on this fact, a similar response can be expected from AI.

Bias In Orthodontics

In orthodontics, bias is defined as the failure to achieve desired goals in treatment. This can be due to improper analysis of the case or selection of inefficient treatment models. Thus, one aim of using AI in orthodontics is to minimize bias and offer the ideal treatment option for the patient.

Multimodal Learning: Recognizing Bias In Orthodontic AI

Orthodontists rely on multiple data sources for accurate diagnosis and treatment planning. Multimodal learning is a field of AI that merges different types of data into 1 model for better prediction of outcomes. The types of data usually combined in the method include:

  • Radiographic images
  • Texts
  • Numerical metrics

The stage in which data is combined and fed into a single model is known as the fusion stage.

The most straightforward approach in data combination is early fusion. Data extracted from lateral cephalograms, clinical notes, and clinical data is converted into a tabular form. As the data set (to be selected) is handpicked, there is incorporation of bias.

The first application of multimodal AI in orthodontics assessed the need for orthognathic surgery. The diagnosis success rate of the model was 96%. In another study, the accuracy of this model’s prediction for determining whether orthodontic extraction is needed or not was 93.9%.

In these cases, model training was achieved by combining manual landmark detection in lateral cephalograms and features (metric) extracted from dental casts. Despite the high accuracy, there was a great level of bias involved.

How Can Bias And Disparities In Orthodontic AI Be Mitigated?

Neural Network Based Data Extraction

To minimize bias, an intermediate fusion was applied. This particular technique combines modality-specific features with the help of an artificial neural network (ANN). So, instead of manual landmark determination, data is extracted by means of an AI-based neural network (which can be from single or multiple input modalities). As new data is fed with every new iteration, the features extracted by a neural network continually change. Plus the data extraction process surpasses human experts, thereby reducing the risk of bias.

Accurate Growth Prediction

The next step to mitigating bias-induced disparities is to improve the orthodontic treatment planning. Artificial intelligence can offer precise and customized care to patients to minimize side effects. Particular patient features such as enzyme deficiencies and bone types can interfere with the treatment plans. For example, bone density determines the speed of tooth movement during treatment.

AI can offer superior treatment plans than humans. Sophisticated AI algorithms can accurately give growth predictions which can in turn allow the provision of tailored orthodontic treatment to the individual. The exact time (and age) for functional treatment in Class 2 malocclusion plays a crucial role in prognosis. An artificial intelligence system can predict growth and provide doctors with the ideal time frame for treatment.

Expanding The Data Set

AI can eliminate the deep-rooted issues of healthcare and medical research with relative ease. As already mentioned, bias in research and clinical trials is the foundation for bias in treatments. Data dentistry can be transformed by using AI’s deep learning. As AI algorithms are capable of handling large amounts of data, we expect artificial intelligence to provide better and more precise solutions (without difficulty).

Experts believe this revolution can be brought by integrating different types of clinical (dental) data into deep learning. Sets of data include:

  • Demographic
  • Social
  • Clinical and omics
  • Consumer
  • Geospatial
  • Environmental

Replication of the machine learning models is a challenge in orthodontics because most of the patient data (which serves as the foundation for learning models) is inaccessible to others due to privacy. However, the quality of reporting can be enhanced with systems such as CONSORT-AI, TRIPOD, etc.

Final Word

AI systems in healthcare rests on the foundation of clinical research. Therefore, bias in AI stems from bias in clinical trials and research. To recognize bias in Artificial intelligence programs we understand how it works. In the initial days of AI, an early fusion model was used for orthodontic analysis. It combined radiographic and clinical data selected by experts (human choice adds bias). To mitigate this, neural-network-based intermediate fusion is utilized to minimize human-based decision-making. Moreover, AI can offer individualized growth predictions that reduce bias. It can also alleviate bias levels by adding multiple types of data sets in research. Demographic, environmental, social and geospatial data integration in case analysis can reduce bias to a minimum.

References

  1. Steinberg, J. R., Turner, B. E., Weeks, B. T., Magnani, C. J., Wong, B. O., Rodriguez, F., … & Cullen, M. R. (2021). Analysis of female enrollment and participant sex by burden of disease in US clinical trials between 2000 and 2020. JAMA Network Open, 4(6), e2113749-e2113749.
  2. Riolo, M. L., Moyers, R. E., & TenHave, T. R. (1988). Imprecision and bias in orthodontic treatment results. American Journal of Orthodontics and Dentofacial Orthopedics, 93(2), 138-142.
  3. Choi, H. I., Jung, S. K., Baek, S. H., Lim, W. H., Ahn, S. J., Yang, I. H., & Kim, T. W. (2019). Artificial intelligent model with neural network machine learning for the diagnosis of orthognathic surgery. Journal of Craniofacial Surgery, 30(7), 1986-1989.
  4. Del Real, A., Del Real, O., Sardina, S., & Oyonarte, R. (2022). Use of automated artificial intelligence to predict the need for orthodontic extractions. Korean Journal of Orthodontics, 52(2), 102-111.
  5. Chugh, T., Jain, A. K., Jaiswal, R. K., Mehrotra, P., & Mehrotra, R. (2013). Bone density and its importance in orthodontics. Journal of oral biology and craniofacial research, 3(2), 92-97.
  6. Perinetti, G., Primožič, J., Franchi, L., & Contardo, L. (2015). Treatment effects of removable functional appliances in pre-pubertal and pubertal Class II patients: a systematic review and meta-analysis of controlled studies. PloS one, 10(10), e0141198.
  7. Schwendicke, F., & Krois, J. (2022). Data dentistry: how data are changing clinical care and research. Journal of dental research, 101(1), 21-29.

You'll find more articles in my blog:

Read more