Potential of artificial intelligence in injury prevention research and practice

Abstract

There is increasing interest and use of artificial Intelligence algorithms and methods in biomedical research and practice, particularly as the technology has made significant advances in the past decade and become more accessible to more disciplines. This editorial briefly reviews this technology and its potential for injury prevention research and practice, proposing ways that it can be used to advance the discipline, as well as the potential pitfalls, concerns and biases that accompany it.

Over the past decade, and especially in recent years, artificial intelligence (AI) has permeated news, politics, and many aspects of everyday life (e.g., chatbots, virtual assistants, social media, smart devices). Biomedical and public health researchers and practitioners are also finding uses for AI. AI algorithms have been used to radiography and biomedical imagery, medical records,[1] and to identify built environment features associated with health outcomes. What potential do they have for injury prevention and control? A brief literature search suggests these methods are also being adopted by the field: examine road infrastructure safety and crashes,[2] predict the severity of motorcyclist injuries,[3] detect motorcycle helmet use,[4] predict and prevent sport injuries,[5] and to identify built environment typologies related to firearm violence.[6] What implications do these have for the field and how can we adopt them along with more traditional approaches?

AI refers to both a set of algorithmic approaches to analysing data, as well as the theoretical underpinnings of the discipline that has the goal of creating or simulating intelligence.[7] AI encompasses both machine learning and deep learning, the latter becoming most synonymous with AI in recent years. Deep learning algorithms underlie natural language processing models, speech recognition used in voice assistants, object detection and recognition in video and images like facial recognition and self-driving vehicles, and generative AI that create different media. AI as a field has existed since the 1940s and 1950s and many of the early successful efforts centred around text and language, though other areas of research included computer vision, speech and audio, and video.[7, 8] The field has passed through cyclical ups and downs, with current interest and progress in AI beginning in the early 2010s when newer models demonstrated high accuracy at or near human-level performance.[7] This has been possible by a combination of computing power available, distributed or cloud computing, algorithmic developments, and graphics processing units (GPUs).[9] The most common sets of algorithms that deep learning rely on are neural networks and diffusion models. While more technical details on these models can be found elsewhere, a brief overview will be provided here.[7, 10–12]

The essential structure of neural networks consists of various processing layers that examine and transform patterns in the data as they pass by each layer. The structure and connection between layers vary by the type of neural network being used with some of the most types being convolutional neural networks (CNNs), recurrent neural networks (RNNs), long short-term memory (LSTM), deep belief networks (DBNs), generative adversarial networks (GANs), and transformer neural networks. Neurons are the essential building blocks of neural networks and are meant to emulate the neurons of the human brain. The neurons are essentially a set of mathematical functions that receive input data that are weighted and summarized to then produce an output via a process known as activation. That output may then be used as input for the neurons in the next layer. Neurons are connected to one other throughout the layers and their combined weights and outputs are then used to produce the final outputs of the model (e.g., objects in images, words in text or speech, etc.). Neural networks also have other characteristics that determine a neural network’s structure (e.g., number of hidden layers and units, dropout, initialization, and activation function) and how the network is trained (e.g., learning rate, momentum, number of epochs, and batch size) known as hyperparameters. Most recent models also integrate backpropagation, which compares prediction outputs from the model with observed data to correct errors in the weights. Models like CNNs rely on feeding data sequentially forward through the network. RNNs also feed data forward and can also feed data backward. Transformer networks, in contrast, can process data in parallel due to their reliance on a self-attention mechanism that allows the model access to all data elements at once to learn the context or relationship of each element to every other element in the model.[13, 14]

Neural networks rely on training data to accomplish the tasks they are designed for and may be semi-supervised or unsupervised in this process. Semi-supervised models rely on training data that has been tagged, labelled or written by humans that can teach the neural network what the objects of interest are. Unsupervised models, on the other hand, are trained on data without human input to learn the objects of interest within the data. Unsupervised models typically need substantially more training data than other neural networks, for example some of the leading transformer networks available today have been trained on billions of data elements or more.[13, 15] In both cases, the models learn underlying characteristics and patterns in the training data through an iterative process of transforming the training data as it passes through each layer of the network. The performance of the models is typically measured through metrics of precision and recall, which are better known to epidemiologists and biostatisticians as the positive and negative predictive values, respectively. A summary metric, mean average precision (mAP), is created from precision and recall for evaluating models that is similarly interpreted as the area under the curve (AUC) metric in receiver operating characteristic (ROC) analyses.

Diffusion models are relative newcomers to AI, but are quickly becoming integral to generative AI enabling the generation of text, images, video, sound, etc. from pre-trained transformers.[16] They originate from the field of physics, particularly the study of non-equilibrium thermodynamics that includes the examination of the dispersion of substances like gases.[17, 18] Essentially, these models randomly add noise to the input data over several steps until all the data are uniform and then the process is reversed. The model learns from this process to create a set of probability densities of the input data, which can then be used to generate novel data.[16]

How can injury researchers and practitioners take advantage of these algorithms? First, these methods present immense opportunities to collect data on a much more massive scale than previously possible. Natural language processing (NLP) models can review records, extract data, and make clinical predictions from electronic health records (HER) more efficiently and quickly than human reviewers.[12, 19] Computer vision models can review millions of images to find relevant objects of interest, such as road traffic environment characteristics from street images,[20] neighbourhood characteristics from satellite imagery,[6] and identifying clinically significant objects in medical imagery.[21] These models can also identify people, vehicles and other moving objects in videos to produce counts that can be predictive of transportation mode.[22] Audio models can transcribe recordings of qualitative interviews, focus groups and other events, which can then be coded by text-focused models and analyzed.[23] These methods can also be used for mining social media to find patterns of interest, such as monitoring trends in mental health, violence, perceptions and sentiments.[24, 25] An emerging area of research is using generative models to produce simulated data to replace real-world data, particularly for protecting the privacy and confidentiality of medical data.[26, 27]

Second, just as machine learning has enabled analytical advances with large datasets, deep learning can further these efforts by replacing or enhancing some of the human effort or statistical methods. These methods can be more efficient for data reduction to create indices,[28] improve multiple imputation for missing data,[29] predicting health outcomes and risk factors.[30, 31] AI can also be used to assist with drafting communication materials,[32] assisting with analytical coding or programming,[33] visualizing changes to road infrastructure or other image and video outputs,[34] presenting and visualizing data,[35] and answering questions about injury prevention for educational purposes via chatbots.[12, 36, 37]

As with any new technology or methods, there are important pitfalls and biases that should be acknowledged and understood when using AI for research. A key challenge for most researchers will be coming to a fundamental understanding of how they work and how to apply them.[12] Injury researchers and professionals should be focused on finding expert collaborators and partners that can implement the desired data collection, analysis, or communication, just as we collaborate with other experts in other areas outside our specific discipline. Another challenge is ensuring the training data used for the models that researchers adapt to their needs are reliable and high quality.[38, 39] Many current models are trained on data from high resource settings in the U.S. and Western Europe and in English language, thus their generalizability to most of the world may not be appropriate. Many models are not entirely open, meaning the training data, the model weights, or other aspects are not available publicly; thus it is unknown how they were trained and validated. This, in part, contributes to “hallucinations” that some models may have, meaning generating text, imagery or video that is incorrect or inaccurate.[37] An additional issue with training data important to researchers is copyright and fair use, which differs from country-to-country.[40–42] Obtaining high quality data or permission to use high quality data can be expensive and challenging, thus researchers may turn to publicly available resources, yet those products may have important terms of use or service that limit use, even fair use in the US legal context.[41] The role of fair use in training artificial intelligence models is yet to be resolved, particularly for generative models.[42] The environmental impact of these models is also an important consideration as the computing power and resources needed to create new models is immense, as well as their use for customizing for a more specific use.[43] Another cause for concern is that they are a black box, even to those that have created them due to a fundamental inability to understand why these models work, how they work, and of why they may fail.[15] For causal epidemiologists and seeking to understand how exposures are related to disease, this could be problematic, particularly if relationships are observed that are unexpected or that cannot be explained by traditional means.[44, 45] Finally, there are many equity issues surrounding AI models and methods, such as the specialized knowledge needed to successfully use them, the access to the computational resources to train and implement them, the data that has been used to train them and access to those data, where they are being employed, who is involved in developing these models, and worsening existing disparities.[46–49] Some researchers have proposed all models should be accompanied with “model cards” or “datasheets” that provide standardized documentation that can make the models more transparent, ensuring they comply with FAIR (findable, accessible, interoperable, and reusable) data principles, and regulating the development and implementation of AI.[48, 50–52] There are also a growing number of recommendations for practitioners and researchers for promoting and ensuring health and healthcare equity with the integration of AI into the healthcare setting that could also be considered for developing similar recommendations for research use.[53–58]

In conclusion, artificial intelligence via deep learning has the potential to improve our efforts as injury prevention researchers, practitioners, and policymakers by improving efficiency in data collection and analysis, as well as increasing our reach and many other uses not described here. Injury Prevention welcomes submissions that examine and validate AI models and approaches in injury research and policy, including the evaluation of their ethical, equitable, and methodological soundness.