The research – published in the journal BioMed Central Medicine – found the new AI model was able to accurately predict the most effective drug for individuals using real patient data. It found that people who were prescribed the medication recommended by the AI model had a slower decline in thinking and memory skills, known as cognitive performance.
The researchers were mostly from the University of Oxford’s Department of Psychiatry, and the team was led by Professor Alejo Nevado-Holgado with Dr Qiang Liu, a postdoctoral researcher within the department, as first author. The paper was co-authored by Dr Ivan Koychev, a Senior Clinical Researcher in the Department of Psychiatry and Research Volunteer Registry Lead at DPUK.
Dr Liu said: ‘We try to leverage AI to identify the best medication for each individual patient diagnosed with dementia based on their own unique patient journey and treatment pathway. We found the patients who followed our model recommendations showed an approximately four-fold slower cognitive decline two years after treatment initiation, when compared with patients who did not.’
Four medications are commonly prescribed to treat dementia in the UK, US and Canada: donepezil, galantamine, rivastigmine and memantine. Usually, the drug prescribed largely depends on cost and personal preference of the clinician, except memantine which tends to be prescribed later in the disease progression.
AI is already used in dementia research to monitor brain health and predict the onset of dementia as early as possible. These tools often use deep neural networks, which are complex algorithms inspired by the structure and function of the human brain.
Dr Liu and his team built their cognitive score prediction model based on these existing deep neural networks. Their new model combines seven variables – age, sex, ethnicity, marital status, time between clinical observations, and cognitive scores – to predict the extent of cognitive decline that would occur with each of the four medications.
The Oxford-based researchers used datasets of 6,804 people with dementia aged between 59 and 102, collected in two UK NHS Foundation Trusts. They developed the AI model using data from 60% of the Southern Health NHS Foundation Trust (SHFT) dataset, before testing the model on the remainder of this dataset as well as the Oxford Health NHS Foundation Trust (OHFT) dataset.
These graphs demonstrate the difference in cognitive decline between the two groups, with graph A reflecting the SHFT dataset and graph B describing the OHFT dataset. NNToC stands for neural network treatment of choice and describes the group who received the medication recommended by the AI model.
The results show that patients in both groups were stable for around six months after first being prescribed medication for their dementia, which is consistent with other studies. The current research found that this was followed by a significant decline in cognitive performance, with people not on their AI-recommended drug declining faster and further. This research provides evidence that personalised drug selection can extend the positive effects of dementia medication beyond the first six months.
Dr Ivan Koychev said: ‘In this study using real-world NHS data we show that tailoring the choice of cognitive enhancers of dementia to the patient can extend the period during which they are effective by up to a third. This is potentially important for those of us working clinically as currently there is very little to guide us in the choice between the four different cognitive enhancer options.’
Interestingly, only 12.53% of people with dementia in the OHFT cohort were prescribed the optimal medication for them, as predicted by the AI model. To combat this, the researchers are currently developing an online version of their AI model, which they hope will be used by doctors to help them decide which medication to prescribe their dementia patients.
To test their algorithm’s accuracy, the researchers grouped patients into those who had received the medication recommended for them by the model, and those who had been prescribed a different medication. They then looked at the actual change in cognitive scores for the patients in each group. A lower cognitive score indicates more severe cognitive impairment – read this blog post for more information about cognition.
Patients who received the drug predicted to be best suited to them showed better cognitive performance for two years after being prescribed their medication, as demonstrated by a slower decline in cognitive scores.