A-Z of AI in Healthcare
Subgroup determination
Drugs often have different effects on different groups of people, both in terms of how the drug affects the patient and the size of that effect. These different groups are known as ‘subgroups’ of the population.
When developing a new drug or new treatment, it’s important to understand which patients respond best to the drug (i.e., have the best results), which patients have a moderate response (i.e., have OK results), and which patients have a negative response (i.e., have bad or dangerous results such as an allergic reaction). These three types of patients are known as ‘subgroups’ because they are smaller groups within an overall population.
Working out which patients belong in which group is a process known as subgroup determination.
AI, particularly unsupervised AI, can help with this process. For example, you might give an unsupervised algorithm an unlabelled dataset containing the electronic health records of all the patients enrolled in a clinical trial including the ‘results’ of the new drug or treatment (e.g., lowering blood pressure). You can then tell the algorithm to group similar patients according to the results (a task known as clustering). In this way, the algorithm can determine the subgroups and assign patients into the different groups.
Using AI to help with subgroup determination like this can be helpful because it might indirectly identify potential ‘causes’ for the different responses amongst patients that were not immediately obvious. It might, for example, identify that all patients that were in the ‘good response’ group were between the ages of 18 and 45, all those with a moderate response were between the ages of 45 and 65, and all those with a poor response were over the age of 65.
This is very useful data because if the drug was to ever be made available to clinicians as a possible treatment for blood pressure, this information might contribute towards possible guidance issued recommending that the drug be used as the primary treatment for younger patients with high blood pressure, but not as the primary treatment for older patients.
Drugs often have different effects on different groups of people, both in terms of how the drug affects the patient and the size of that effect. These different groups are known as ‘subgroups’ of the population.
Identifying or determining these subgroups and the variations in their treatment response is a crucial step in drug discovery, drug safety monitoring, and the development of personalised medicine. This can be done both prospectively (i.e., ahead of a clinical trial) and post-hoc (i.e., after the trial has started and during the real-world testing of the drug).
Prospective subgroup identification (known formally as confirmatory subgroup analysis) involves the identification of a small number of predefined covariates (typically demographic patient characteristics known as biomarkers) that are listed in the registered trial protocol. For example, the COVID-19 vaccine trials, recruited trial participants and tested the various COVID-19 vaccines, in different age groups.
Post-hoc or ‘exploratory subgroup analysis’ is a more data-driven process that increasingly relies on the use of machine learning techniques (such as the use of unsupervised clustering algorithms) to discover new subgroups based on the analysis of a very large number of covariates (i.e., demographic, genomic, clinical, or other patient characteristics) and their impact on treatment response. This is an important process, not only for identifying which ‘patients’ may benefit the most from a specific drug, but also for the prevention of overfitting and bias, and making the estimates of treatment effect size more ‘honest’.
There is no one agreed method for conducting post-hoc exploratory subgroup analysis, and whilst there are guidelines available to guide the process of prospective confirmatory subgroup analysis no such guidelines yet exist for data-driven subgroup discovery. It is likely that such guidelines will develop over time, particularly as the drive towards personalised or precision medicine gathers pace. In the meantime, it is important that those conducting post-hoc analysis, particularly those using machine learning techniques, document the process, pay attention to basic best practice guidelines for statistical analysis (e.g., how to avoid p-hacking), and consider any appropriate responsible AI guidelines.
