“Data is dynamic and changes its statistical features over time. Therefore, AI models need context-specific and updated data on a frequent basis to ensure accuracy and maintain trust.” This highlights the need for updated AI systems in the healthcare system as described in the “Artificial Intelligence for Strengthening Healthcare Systems in Low and middle-income Countries” review.
Relying only on a static set of data to train AI models is insufficient in the constantly evolving world we operate. The calibre of data is the determining factor in the strength of algorithms and reliability of predictions, regardless of the setting. This reality is especially impactful in the healthcare technology industry where trust is paramount, and a dependable AI system that performs well could mean the difference between life and death.
The potential for AI to revolutionize healthcare is tremendous. AI-driven solutions can enhance the accuracy of disease diagnosis, treatment, and prevention, as well as optimize administrative duties and increase the efficacy of healthcare systems. However, there are several obstacles that need to be addressed before AI can be fully integrated into the healthcare sector. These challenges are:
- Transparency: Trust is crucial, and transparency in AI algorithms is essential to ensure accuracy and reliability. The meticulous selection of data used to train these AI algorithms is equally important.
- Usability: User-friendliness and comprehensibility of AI-powered tools are imperative to ensure wide adoption in healthcare settings. Therefore, designing tools with usability in mind and clear instructions is necessary.
- Affordability: One of the major challenges faced by AI-powered tools is the cost involved in development and deployment. This expense can pose a significant hurdle for low- and middle-income countries (LMICs) to adopt such tools.
Addressing these challenges requires adopting innovative and scalable solutions which will help optimize this technology for healthcare delivery and the ultimate goal of global health equity. To actualise this, it’s important to focus on the following;
Transparent algorithms: In the world of Artificial Intelligence, transparency is of the essence. The utilization of the ‘black box’ technique, which lacks clarity on the algorithms’ final outcomes, is becoming more widely recognized as an impediment to gaining trust. To establish trust among users and stakeholders, it is essential for an AI system to provide transparency regarding the source of data, processing techniques, and algorithms employed. This fosters an environment where users can understand, validate, and therefore trust the outputs provided by the AI system.
In the published review of adopting AI in strengthening healthcare systems in LMICs, one major challenge identified was the incompatibility of the AI system with existing electronic health record systems. This made it difficult for the AI tool to support decisions with relevant health record data, affecting the tool’s performance. Greater transparency and availability of the algorithms and datasets used in training the AI would enable a deeper understanding of its performance, thereby increasing trust in its usability.
Bridging the gap between physician decision-making process Vs AI’s decision
In the context of healthcare, bridging the gap between physician decision-making and AI’s decision is a complex yet crucial endeavour. On one hand, medical professionals utilize their extensive education, practical knowledge, and inherent comprehension of the patient’s individual situation to make informed decisions. On the other hand, AI systems rely on data input, machine learning algorithms, and statistical analysis to reach conclusions.
The ideal AI system should not seek to replace the physician but rather, complement their expertise. By offering a transparent analysis of vast amounts of patient data, AI can help physicians make more informed decisions, in less time. However, to build trust, AI systems should be explainable, and be able to qualify AI-based suggestions when used by clinicians to make decisions.
Adopting uncertainty estimation methods
In the field of artificial intelligence, the ability to accurately predict outcomes is a key element of their utility. Notwithstanding, in light of the intricate and occasionally unforeseeable character of healthcare information, prognostications produced by AI mechanisms may possess a certain degree of ambiguity linked with them. The ability to quantify this uncertainty can be invaluable to physicians and healthcare workers making critical health decisions.
Uncertainty estimation methods, like Bayesian Neural Networks and Monte Carlo Dropout, provide ways to incorporate and quantify this uncertainty.
Bayesian Neural Networks (BNNs) are a type of neural network that uses Bayesian probability theory to estimate uncertainty. The weights within a BNN are not a single value but rather a distribution over potential values. This allows the network to express doubt about its weights and predictions. Throughout training, the distributions are adjusted to reflect the data, resulting in a network that can make probabilistic predictions.
Monte Carlo Dropout is a method used to approximate Bayesian inference in deep Gaussian processes, which is a type of machine learning model. This technique involves applying dropout during both training and testing. Dropout is used to prevent overfitting during training, and the model can estimate the uncertainty of its predictions by performing multiple forward passes through the network and averaging the results.
Both of these techniques provide a means of estimating the confidence of an AI system’s predictions. In healthcare, this additional information can be critical in the decision-making process, allowing physicians to assess the certainty of an AI system’s prediction when making clinical decisions.
Incorporating edge computing
Edge computing refers to the paradigm where computation is moved closer to the source of data generation, reducing the need for long-distance communications between client and server, thereby reducing latency and bandwidth usage. In the context of healthcare, this can provide numerous benefits.
In low-resource settings, such as LMICs, for instance, edge computing can facilitate the utilization of AI applications in real-time, even with intermittent connectivity, eliminating the requirement for continuous internet connectivity. This will allow the processing of patient data on-site, enabling immediate insights and decision-making support to be provided.
Also, edge computing can also bolster data security and privacy. By processing sensitive health data locally, the need to transmit this data over networks is reduced, minimizing the risk of data breaches.
From a cost perspective, edge computing can also make the deployment of AI solutions more affordable. By minimizing the need for data transmission and centralized processing, the overall operational costs can be significantly reduced. This approach can prove advantageous in settings with limited resources where the expenses associated with data transmission and storage might be a hindrance.
Furthermore, edge computing enables healthcare delivery that is more tailored to individual needs. AI models have the capability to be trained on locally sourced data, thereby guaranteeing customization to cater to the requirements of a particular demographic. This, in turn, elevates the efficiency and utility of such models.
Utilizing transfer learning techniques
Transfer learning proves to be a powerful tool in reducing computational resources required for AI solutions. The process involves fine-tuning pre-trained AI models instead of training them from the initial stages to suit a specific task.
In the healthcare industry, a pre-existing AI model can be personalized to evaluate medical images, anticipate the advancement of illnesses, and modify treatment strategies for individuals. This methodology diminishes the amount of time and computational assets required, hence allowing for more affordable and accessible AI usage, even in settings with limited resources.
The journey to AI adoption in healthcare is fraught with challenges. However, the potential rewards – improved health outcomes, greater efficiency, and broader access to healthcare are well worth the effort.
Pivoting towards global health equity
With the promise of transformative potential, AI-driven healthcare solutions can be instrumental in progressing towards global health equity.
- Trust forms the bedrock of any health intervention, more so in AI-enabled healthcare. Transparent algorithms, uncertainty estimation methods, and a strong focus on ethical considerations are essential in building this trust. For instance, Bayesian Neural Networks and Monte Carlo dropout approaches, as discussed earlier, give clinicians insight into the confidence of an AI system’s predictions, improving the trustworthiness of these systems. AI developers must adhere to ethical guidelines to ensure that the systems they create are equitable, non-discriminatory, and respect user privacy. This is crucial in establishing trust in these technologies.
- Affordability is a crucial factor to consider in advancing health equity. Many AI systems require substantial computational resources, which can be expensive. However, strategies such as edge computing and transfer learning can reduce these costs. Edge computing can significantly lower data transmission and storage costs by processing data at or near the source, making AI more accessible in low-resource settings. Transfer learning can leverage pre-trained models to reduce the computational resources needed to develop AI solutions.
- Usability is one of the fundamental pillars that ensure the positive impact of Artificial Intelligence on global health equity. AI solutions should be designed to be easily understandable, user-friendly, and seamlessly integrated into the existing workflows. For instance, integrating AI tools directly into electronic health records allows physicians to benefit from decision-making support while performing their daily tasks. Additionally, AI solutions trained on local data and adapted to the specific needs of the population can significantly enhance their effectiveness.
The convergence of trust, affordability, and usability is the key to the potential impact of AI on global health equity. AI can assist healthcare professionals worldwide, especially in LMICs with limited resources, by offering dependable and cost-effective decision-making aids. When equipped with AI tools that are reliable, affordable, and user-friendly, healthcare professionals can provide enhanced and more informed care, regardless of their location. This brings us closer to a future where quality healthcare is a universal reality, not a privilege.
Source – https://www.nature.com/articles/s41746-022-00700-y
