A DATA-DRIVEN PRIORITIZATION ALGORITHM FOR MEDICATION REVIEW IN OLDER ADULTS: FROM REAL-WORLD EVIDENCE TO CLINICAL PRACTICE

Inappropriate prescribing in older adults represents a growing challenge for healthcare systems worldwide, given the rising multimorbidity and polypharmacy in aging populations. Inappropriate pharmacological regimens not only increase the risk of adverse drug reactions and hospitalizations but also contribute to excessive healthcare costs and a decline in patient quality of life. Given their unique clinical complexity and physiological vulnerability, older adults are particularly exposed to the risks associated with inappropriate medication use, making them a critical target for preventive strategies aimed at optimizing pharmacotherapy and reducing avoidable harm. Among these measures, periodic implementation of medication review is recommended. However, healthcare professionals- particularly general practitioners -are central to the management of chronic therapies, face increasing time constraints often hinder the systematic implementation of medication reviews. In this context, there is a pressing need for flexible, data-driven tools that can support clinicians in identifying patients most likely to benefit from medication review interventions, thereby optimizing the allocation of time and resources. This thesis presents the development of a prioritization algorithm aimed at prioritizing older patients expected to gain the greatest benefit from medication review interventions, based on comprehensive real-world data. Conducted over a three-year doctoral research project, the work addresses a pressing public health concern- polypharmacy and inappropriate drug use among the elderly- and proposes a pragmatic, data-driven solution to enhance prescribing safety and optimize clinical resources in primary care. After in depth literature review, six prescribing inappropriateness indicators (PIIs) were identified: • Potentially inappropriate medications (PIMs); • Potential drug–drug interactions (pDDIs); • Fall-risk-increasing drugs (FRIDs); • Therapeutic duplications (TDs); • High anticholinergic cognitive burden (ACB); • High sedative load (SL). The identification of these indicators was based on established international criteria, adapted to the Italian drug market and validated through robust operational definitions. The work was conducted using health care administrative databases from four Local Health Units in Lombardy (Italy), covering over 640,000 individuals aged 65 years or older between 2015 and 2018. Exposure to each PII was determined by operational definitions accounting for drug dispensation dates, ATC classification, Defined Daily Doses and co-prescribing windows. Individuals with first exposure between January 2015 and September 2018 were followed for 90 days to evaluate the risk of all-cause hospitalization, using a risk-set matching strategy with one control per exposed individual, matched on age and sex. Multivariable logistic regression models, adjusted for high-dimensional propensity scores, were used to estimate the odds ratios (ORs) of all-cause hospitalization associated with each PII. All six PIIs were significantly associated with increased hospitalization risk. The highest risks were observed for FRIDs (OR 1.62, 95% CI: 1.51–1.73), PIMs (OR 1.55, 95% CI: 1.48–1.62), and TDs (OR 1.52, 95% CI: 1.46–1.58). SL score (OR 1.43, 95% CI: 1.31–1.56) and ACB score (OR 1.52, 95% CI: 1.43–1.61) Established the baseline OR for each PII, we developed decision tree models for each indicator to identify high-risk patient profiles- i.e., combinations of demographic and clinical characteristics associated with significantly elevated hospitalization risk beyond that conferred by exposure alone. Each scenario was validated using conditional logistic regression in comparison with the overall risk of the indicators alone. Across multiple bootstrap samples, 59 distinct high-risk scenarios were identified: 14 for SL, 14 for TDs, 11 for FRIDs, 10 for pDDIs, 7 for PIMs, and 4 for ACB. These scenarios characterize subgroups of patients who should be rapidly prioritized for clinical attention, as they are at particularly high risk of hospitalization. To estimate the real-world impact of using the algorithm, we applied the scenarios to a cohort of 501,785 older adults in the first half of 2018. A total of 55.3% (n=270,487) of individuals had at least one PII. Among these, 11.2% matched at least one high-risk scenario, representing 20.2% of those with inappropriate exposure. SL score and PIMs exhibited lower prioritization efficiency (4.5% and 3.5% respectively). Conversely, pDDIs and TDs flagged over one-third of exposed individuals as high-risk. By considering prioritizing all individuals with at least three indicators, as well as those with one or two indicators who fell into the highest-risk scenarios, a total of 49–52 patients per physician would be selected. This work presents a novel and pragmatic framework to support prioritization in medication review processes. By leveraging routinely collected administrative data, the proposed algorithm enables targeted identification of older adults at heightened risk for adverse drug events. The model’s integration of individual-level risk profiles and cumulative burden metrics offers a powerful and scalable tool for supporting deprescribing strategies in real-world clinical practice, particularly within constrained primary care settings. This work lays the foundation for future prospective validation studies and the potential integration of the algorithm into digital clinical decision support systems.