Clinical trial data reconciliation is a critical process to ensure the accuracy, consistency, and regulatory compliance of trial datasets. Among the most complex reconciliation tasks is the alignment of adverse event (AE) data with concomitant medication (CM) records. This task is prone to inconsistencies, delays, and operational bottlenecks that can impact data quality, regulatory submissions, and patient safety evaluations.
Thankfully, AI is here to help.
Reconciling Adverse Events and Concomitant Medications
One of the most challenging tasks in reconciliation is resolving adverse event (AE) data with concomitant medication (CM) records. These datasets are typically captured independently—AEs by site investigators during clinical assessments and CMs during medication history or ongoing treatment logs. Ensuring consistency between the two requires careful cross-checking of timing, indication, and treatment rationale – a significant bottleneck of human effort.
For example, if a patient reports a severe headache (recorded as an AE), and later receives acetaminophen (recorded in CM), there must be a clear temporal and therapeutic link between the two. If such links are missing or contradictory—such as a serious AE with no corresponding treatment, or a potent medication administered with no AE justification—it raises red flags during data review and regulatory inspection.
Further complications arise when medications are recorded for prophylaxis, pre-existing conditions, or protocol-mandated therapies, making it difficult to determine if they are related to reported AEs. Reconciling these nuances requires medical review, precise coding (e.g., MedDRA for AEs, WHO Drug Dictionary for CMs), and often manual adjudication to confirm alignment.
To support this process, Readout’s AI performs a number of these tasks. First, we can code adverse events to MedDRA and medications to WhoDRUG, all automatically. Our AI can even put an Adverse Event into the most reasonable category if it doesn’t exist in the MedDRA dictionary.
But more importantly, our AI has been developed to reason about whether it would be unusual for an AE to not be associated with a medication (versus resolving itself) and whether it would be unusual for a CM to be recorded without an associated AE (e.g., maybe it was prophylactic or a day-to-day medicine).
Two other considerations for this model are that it’s “stable” (the AI produces the same results across runs) and it has a sense of uncertainty. When it’s confident that it’s flagged an issue (or there is not an issue), it will say so, but when it’s not, it will also alert the user.
The goal is for the AI to assess whether further, human review might be required given that an adverse event was recorded (with no associated concomitant medication to address it), or vice versa (that a medication was recorded without seemingly being prompted by an adverse event). If the AI can help screen out the cases where a human doesn’t have to put in effort to review (e.g., a “true negative”) then the AI is saving significant human effort.
Here is one such case. In this case, it’s obvious that no medication needed to be recorded (e.g., associated with this adverse event), and the AI explains why.
Pyrexia (fever) with a severity of mild and an outcome of recovered/resolved is often managed without specific medication, especially if it is a low-grade fever and the patient is otherwise stable. Supportive care, such as rest and hydration, is frequently sufficient. Considering the patient’s medical history includes conditions like Alzheimer’s disease and other chronic ailments, and given the concurrent mild adverse events of dehydration and hypotension, a decision to monitor and manage the fever without immediate medication is not unusual. The previous concomitant medications (Vitamin E, Bactrim, Keflex, Ecotrin) do not directly suggest a need for immediate treatment of the fever. The outcome was recovered/resolved which suggests medical intervention was not needed.
That is, the human can give this explanation a quick review and no that no reconciliation action is needed.
On the flip side, the AI can also help identify those cases that do require more human analysis. Here is one real example where the AI is confident that a medication should likely have been recorded for this adverse event and therefore a person should double check the reconciliation. The patient has a significant medical history (10+ issues in the medical history, including Alzheimer’s, Hypothyroidism, and Osteoporosis) and has recorded a number of concomitant medications prior to recording an Adverse Event of moderate cellulitis. The AI believes there might be a missing concomitant medication recording, reasoning that:
Cellulitis is a bacterial skin infection that typically requires antibiotic treatment. Given the moderate severity of the cellulitis and the fact that the outcome was reported as ‘Not Recovered/Not Resolved,’ it would be unusual not to treat it with medications, specifically antibiotics. While some minor skin infections might resolve on their own, cellulitis, especially when moderate, generally necessitates medical intervention to prevent complications and promote healing. The patient has a history of other infections (Bladder Infection, Tuberculosis, Localized Infection), and while taking a previous antibiotic (KEFLEX). The cellulitis arose despite that, which might imply a resistance of some kind or different bacteria entirely.
In other words, the AI has ruled out that the previously administered antibiotics treated the Cellulitis successfully (according to the data) and the patient has a history of infection – so it’s likely that a medication should have been recorded to treat this event (and is it unlikely that they simply left it untreated).
Finally, here is an example where the AI is less certain – it thinks there “might” be an issue, and therefore still warrants further human analysis, as it can’t make a strong judgement.
Sinus bradycardia, even when mild, can potentially lead to more serious cardiovascular issues if left unaddressed. While some cases may resolve on their own, it is a condition that is often treated with medication, depending on the heart rate and the patient’s symptoms. Medications like atropine or other anticholinergic drugs can be used to increase the heart rate. It would be unusual to not consider medication as a treatment option, especially given the patient’s existing medical history including hypertension and systolic murmur, as untreated bradycardia could exacerbate these conditions or lead to complications.
This is an interesting case because, as noted by the AI, the decision to treat medically depends on some specific symptoms, which we don’t have at this point in the trial (hence the need to go back to the site for details in the reconciliation).
So, the AI saves time by weeding out the more obvious cases where no reconciliation is required while helping improve data quality by surfacing the cases where it is. And, so as to be as conservative as it can, it also surfaces the cases where it’s not sure.
Reconciliation for AE-CM alignment is essential but fraught with operational and analytical complexities. But with AI support and strong collaboration between clinical, data management, and safety teams these challenges can be addressed to ensure high-quality, audit-ready clinical trial data.
