Big Data/Machine Learning/AI
Leveraging Nonparametric Machine Learning to Assess Effect Modification of Cervical Precancer Treatment in Women Living with HIV Nazha Diwan* Nazha Diwan Michael H. Chung Sharon A. Greene Judith Lukorito Samah R. Sakr Nelly R. Mugo Gabriel Conzuelo Rodriguez
Background: Ablation is currently a standard treatment for cervical precancer (CIN2+) in low-resource settings. While its efficacy in non-HIV populations is established, evidence suggests suboptimal effects in women living with HIV (WLWH). It is hypothesized that the elevated risk of recurrence among WLWH depends on the level of immunosuppression. However, most studies assessing this treatment effect modification rely on subgroup analyses of categorized CD4 count, which results in substantial loss of information and statistical power. Flexible parametric models address these limitations by avoiding categorization of continuous modifiers, but still rely on restrictive assumptions about the data-generating mechanism. Alternatively, nonparametric machine learning approaches (e.g. DR-learner) avoid parametric assumptions and provide flexibility to capture effect modification.
Methods: We used data from a randomized trial conducted at the Coptic Hope Center in Nairobi, Kenya (2011-2016), which assessed differences in CIN2+ recurrence over 24 months in WLWH assigned to ablation (cryotherapy [n = 172]) or excision (LEEP [n = 167)]. Effect modification by CD4 count at treatment was modeled using generalized linear models that incorporated either restricted cubic splines or fractional polynomials, as well as the DR-learner.
Results: Overall, LEEP is superior to cryotherapy in treating CIN2+ lesions among WLWH. However, as CD4 count increased, the risk difference for CIN2+ recurrence between cryotherapy and LEEP decreased.
Conclusion: While excisional treatments like LEEP are superior, they are impractical in low-resource settings. Ablation is a viable alternative for some WLWH, with eligibility guided by immunological factors. The DR-learner could be leveraged to more effectively model complex relationships between factors such as ART duration, viral load, and HPV type, enhancing risk stratification and enabling more targeted and evidence-based treatment strategies for WLWH.
