Artificial Intelligence and Clinical Implementation

What are the barriers to algorithm adoption among physicians and healthcare facilities and how can they be overcome?

We all know the progress and speed with which Artificial Intelligence (AI) is entering different sectors. Thanks especially to deep neural networks such as Machine Learning and Deep Learning, algorithms to support healthcare practice, i.e., public health, are being commercialised.

Some of the prevalent methods used for AI systems are as follows:

• Machine Learning: the computing method to 'learn' or derive rules or patterns without the latter being explicitly programmed. This can be supervised machine learning (ML) if it learns from data that includes correct answers, or unsupervised ML if it finds clusters of similarity, and reinforcement learning if the system learns through trial and error.
• Genetic algorithms: a genetic algorithm is a heuristic search inspired by Charles Darwin's theory of This algorithm reflects the process of natural selection where the fittest individuals are selected for reproduction in order to propagate offspring of the next generation.
• Deep Learning: considered a subset of machine learning where algorithms tend to transform data over a series of steps (aka layers). For instance, in computer vision, an initial transformation of a picture is often the result of edge detection (using matrix convolutions) to find boundaries of different brightness within a picture to facilitate the subsequent identification of shapes.
• Robotic Process Automation (RPA): these are software tools that automate a specific procedure and are used to reduce human intervention in purely rule-based processes,g. issuing receipts at the end of each month and archiving them. RPAs can be assisted or unassisted. Assisted RPAs are traditional systems in which a specific set of rules is established. Unassisted RPAs use AI to go beyond the setting of rules.
• Reinforcement Learning (RL): is an area of machine learning concerned with how intelligent agents ought to take actions in an environment in order to maximize the notion of cumulative reward. Its focus is on finding a balance between exploration (of uncharted territory) and exploitation (of current knowledge).

Despite many steps forward, only a few AI-based tools have actually been implemented in the various healthcare systems. Let’s look at the main reasons for this, transparency, data quality, and the ability of doctors and patients to rely on algorithms.

The quality of the data source and thus of the data itself is one of the main concerns of those in healthcare environments dealing with technology. Indeed, it is not always possible to establish the quality of the data and access to the algorithm source code.

In addition, there are still not enough published studies to directly support algorithms that have been tested in silico and may not reflect clinical practice.

How do you establish data quality? In the tech world it is often said: garbage in, garbage out.

Good quality, pre-processed data is even more important than the most powerful algorithms, to the extent that machine learning models trained with bad data could be harmful to the analysis you’re trying to do, giving you “garbage” results.

Raw, real-world data in the form of text, images, video, etc., is messy. Not only can it contain errors and inconsistencies, but it is also often incomplete and doesn’t have a regular, uniform design.

The most common steps in a pre-processing pipeline are:

• Mismatched data types: When you collect data from multiple sources, it may come in different formats.
• Mixed data values: Perhaps different sources use different descriptors for features – man or male, for example.
• Outlier strategy: Outliers can have a huge impact on data analysis results. There are different strategies to handle the outliers, like removing the occurrences or imputing the value.
• Missing data: There are several ways to correct for missing data, but the two most common are ignoring the tuples or imputing missing data with the mean (or median) for numerical values, most present for categorical values, interpolation, and rolling mean in time series.
• Binning: If the data are noisy, you can use binning. It sorts data of a wide set into smaller groups of more similar data.
• Standardization: standardization (or normalization) scales data into a normalized range so you can compare it more accurately.
• Combining features: creates a new variable from two or more features.
• Feature selection: Feature selection is the process of deciding which variables (features, characteristics, categories, etc.) are most important to your analysis.

Even doctors who use AI-based systems to support diagnosis and decision-making rarely have oversight of the data used to reach a specific diagnosis or suggestion, contributing to what is called algorithm aversion.

Algorithm aversion can be mitigated by working on the transparency of the data and the algorithms themselves. For example, providing more transparency about the datasets used for initial training. Or providing more information about the construction of the algorithms in a language that is more understandable to clinicians, and perhaps allowing, at local level, the possibility of integrating the dataset with patient data to help train the neural network.

More progress will be made when open-source AI solutions based on open data become available. These applications would be directly available to clinicians and healthcare providers who could help to build the algorithms with direct patient evidence.

“The medicine of the future is not just a question of devices, hardware, and software. Instead, it will be determined above all by the ways in which people – patients, doctors, health professionals or carers – use and interact with these technologies, and how their habits and mental shape will change as a result.” – Roberto Ascione, Healthware Group CEO.

In conclusion, it will be necessary for physicians to embrace AI for substantial implementation supported by clinical validation and market authorisations.

Additional resources:

• The practical implementation of artificial intelligence technologies in medicine
  https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6995276/
• Clinical Implementation of Artificial Intelligence in Radiology
  https://www.itnonline.com/article/clinical-implementation-artificial-intelligence-radiology
• Duke Center For Artificial Intelligence In Radiology
  https://radiology.duke.edu/research/centers-and-shared-resources/duke-center-for-artificial-intelligence-in-radiology/
• Bridging the implementation gap of machine learning in healthcare
  https://innovations.bmj.com/content/6/2/45
• Role of Artificial Intelligence Applications in Real-Life Clinical Practice: Systematic Review
  https://www.jmir.org/2021/4/e25759/
• Bridging the chasm between AI and clinical implementation
  https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(22)00235-5/fulltext

• Artificial Intelligence
• Autonomous Agents
• Big Data
• Digital health
• Digital health revolution
• Digital Health Transformation
• Education
• Innovation
• Technology

As Head of Technology, Lino leads international teams and assists clients in identifying the most appropriate software architecture and technology solutions for corporate communications, campaigns, service portals, intranet, mobile-enabled websites and digital therapeutics. A strong advocate of R&D and innovation, Lino uses Agile methodologies such as Scrum and Shape Up to get results and ensure projects are delivered to the highest level of quality. With his expertise in AI and DTx, Lino is always at the forefront of emerging technologies and is committed to providing clients with the most effective and cutting-edge solutions.

facebook twitter linkedin