This article addresses a problem of interest to pharmaceutical companies and prescribing doctors, namely how to assess the conditions to change a patient’s medication and how to determine what drug, if any, to substitute.
Please get in touch if you are interested in investing or seek to license this for further development.
In keeping with my approach to AI’s impact, we need to consider how AI offers decision makers new ways of thinking about the challenges we face. My model is post-agentic.
For medicines, I suggest AI will shift our focus from the drug to the drug-patient context. In that respect, the paradigm shift I am describing is that medicines become dynamic variables, not just a pill, and drug substitution becomes a forecast. Therefore, dynamic models of care are more important that static episodic models.
By extension, the value of medicines as assessed by HTA regimes and economic analyses is not evidence that a drug ‘works’, but that the drug ‘changes the patient’s predicted recovery trajectory in a favourable and explainable way.’
I outline here a GDPSS, a general purpose drug substitution system, to frame the paradigm shift from static drug models to predictive clinical science. As healthcare moves towards predictive monitoring and digital twins, drug substitution should increasingly be understood as a simulation problem.
The Problem is Static Medication Thinking
My normal starting point when looking at healthcare challenges is to determine what the right problem is to solve. Much heat and investment is spent on solving the wrong problem.
In medication reviews the questions that dominate are at least these:
- Is this drug indicated?
- Is the dose appropriate?
- Are there known interactions?
- Is the substitute therapeutically equivalent?
- Is it on formulary?
- Is it cost-effective?
These are valid questions, but they do not address the predictive question which is what is likely to happen to this patient if the substitution is made? As soon as you ask this question, all the others look like they are solving the wrong parts of a bigger problem, and in many cases provide no significant improvement.
It is typical and well-established that a medication review is retrospective. It identifies known risks, known interactions, known contraindications, and known prescribing rules. However, the patient’s future clinical state is not static, but we lack ways to meaningfully probe a dynamic future despite evidence that clinical prediction models can problem the near term, 24 to 48 hours in the future, to a high degree of probability (AUC>0.9).
All patients are complex challenges of varying degrees for whom medicines’ efficacy depend on the patient’s physiology, co-morbidity, polypharmacy, functional status, and adherence behaviours. A drug substitution may reduce one risk trajectory while amplifying another; this is consistent with how complex adaptive systems function and in which wicked problems lie waiting to cause exacerbations, adverse drug reactions and medical errors.
At its simplest, substituting one drug for another changes a patient’s future risk state. It may alter cognition, mobility, blood pressure, sleep, pain, hydration, respiratory function, adherence, or the probability of an adverse event.
Medicines are Dynamic Risk Variables
In a post-agentic predictive clinical intelligence system, medicines are modelled. A sedative is not only a sedative; it is a potential modifier of cognition, sleep architecture, night-time mobility, postural stability, reaction time, and falls risk. An antihypertensive is not only an antihypertensive; it may influence orthostatic vulnerability, renal function, dizziness, and blood pressure variability. An analgesic is not only a pain medicine; it may alter sedation, mobility, constipation, delirium risk, respiratory drive, and functional recovery.
My thinking is compatible with familiar clinical methodologies such as drug exposure changes over time, interacts with evolving patient physiology, and alters future risk trajectories. In GDPSS, the medicine is a modifiable variable in a patient-specific forecast.
The General-Purpose Drug Substitution Simulator solves the right problem
The GDPSS is designed to enable clinicians to model the likely clinical and risk consequences of switching a patient from one medicine to another. It is a predictive modelling tool that asks what is likely to change when a medication is added, removed, reduced, increased, or substituted. At its simplest, the model compares a current medication state with a proposed medication state.
The model is asking whether substitution changes the future risk state of the patient. And that is the right problem to solve. This allows drug substitution to be treated as a dynamic clinical prediction problem rather than static judgement.
To be clear, structured drug-response simulation already exists in physiologically-based pharmacokinetic modelling, and other areas, in-silico trials, and digital twins. GDPSS applies this logic to a neglected practical clinical decision: what happens to this patient’s future risk trajectory when one medicine is substituted for another?
Why This Should Matter to Pharmaceutical Companies and Prescribing Doctors
For pharmaceutical companies and clinician, this matters because much of real-world prescribing is substitutional. Patients are switched because of side effects, tolerability, adherence, formulary restrictions, cost, response failure, co-morbidity, or clinical deterioration. Yet the consequences of switching are often poorly modelled. A company able to show where its product improves the predicted risk trajectory relative to alternatives may have a more sophisticated value story than one relying only on population-level trial averages. And for clinicians, they reduce the known risk of medical errors arising from adverse drug events.
This new capability can benefit many stakeholder groups:
- In Market access and HTA evidence generation a substitution simulator could help show how a medicine performs against alternatives in defined patient subgroups, especially where adverse-event avoidance, functional outcomes, or reduced downstream risk matter.
- Real-world evidence strategy using post-market data could be used to refine substitution models and identify where drug switching leads to better or worse clinical trajectories.
- Safer de-prescribing and switching pathways emerge in which patients are not simply started on medicines; they are moved between medicines with modelling the switch itself as important as modelling the target drug.
A Different Kind of Pharmaceutical Evidence
One of the most interesting implications is that substitution simulation creates a different kind of pharmaceutical evidence.
Clinical trials tell us what happened in selected populations under defined conditions. Real-world evidence tells us what happens in practice. A drug substitution simulator sits between these domains. It uses existing evidence, patient-specific data, and dynamic risk modelling to estimate what may happen next in an individual patient or defined subgroup.
The future value of a medicine may depend on how well its use can be predicted, personalised, monitored, and explained in patients. Medicines that can be integrated into such predictive systems may have an advantage. They can be positioned as part of a more intelligent therapeutic pathway so just a product.
This could be hugely relevant to how clinical decision support partnerships between pharmaceutical companies, payers and clinicians emerge and where the objective is to demonstrate medicines can be integrated into predictive, patient-specific decision environments rather than treated as isolated products. Interestingly, for payers, this means that drugs are not items for procurement and paid for a pill by pill basis, but can be priced on robust performative efficacy.
Illustrative use case: patient falls
FRIDs [Fall Risk Increasing Drugs] are a natural first use case because they expose the limits of static medication logic.
A conventional FRIDs review can identify medicines associated with increased fall risk. That is useful, but it is not the same as modelling what happens when treatment is changed. Stopping or substituting a FRID may reduce sedation, dizziness, hypotension, or confusion. But it may also worsen pain, sleep, agitation, mobility, blood pressure control, or behavioural symptoms. In frail patients, the substitution itself may create a new risk state.
Then the key clinical question becomes: if this patient is switched from this medication to an alternative, what happens to their predicted falls risk over the next 24–48 hours and over the following stabilisation period?
Consider an older adult in a care setting who is taking a sedative hypnotic, an antihypertensive, and intermittent opioid analgesia. The patient has mild cognitive impairment, variable blood pressure, reduced mobility, and recent night-time wandering. A static medication review may identify the sedative and opioid as FRIDs. A predictive substitution model asks a more useful question: which change produces the safest future trajectory?
Then model scenarios:
- continue current regimen
- reduce sedative dose
- substitute sedative with lower-risk sleep intervention
- reduce opioid exposure while preserving pain control
- adjust antihypertensive timing
- combine medication change with mobility monitoring.
The prediction results may show that reducing sedation lowers one pathway of falls risk, but destabilising pain control increases another. It may show that changing antihypertensive timing reduces orthostatic vulnerability but requires blood pressure monitoring. It may show that the safest option is not a single drug substitution, but a combined medication and monitoring pathway.
The same predictive approach can be applied to a host of other areas where drug substitution introduces material patient risk, including: antipsychotic switching, analgesic substitution, antihypertensive adjustment, diabetes medicines, COPD therapies, anticoagulation decisions, oncology tolerability pathways, and de-prescribing, especially in frail adults. In each case, the relevant outcome changes, but the modelling question is similar: how does changing the medication state alter the patient’s predicted future risk?
Visually, it looks something like this from a dashboard simulator perspective:
Why This Matters Now
Healthcare is moving towards continuous monitoring, remote care, digital twins, predictive analytics, and AI-assisted clinical decision-making. At the same time, patients are becoming older, more complex, more medicated, and more likely to experience adverse outcomes from interactions between treatment and vulnerability.
Solutions today should be seen as framed in an increasingly dated paradigm. With AI we now ask: what is the predicted consequence of this medication change in this patient? For pharmaceutical companies, this creates a different way to think about product value. The next generation of evidence will not be limited to demonstrating average efficacy in trial populations.
It does, however, support structured clinical reasoning by estimating comparative risk, identifying trade-offs, and making uncertainty explicit. The clinician remains responsible for judgement, and patient preference.
As healthcare moves towards a new paradigm of patient care, we can be reasonably confident that when the current fog clears, predictive clinical intelligence will be the dynamic driving new models of care.