Characterizing the ADPKD- Phenotypic Signature With Deep Learning and Advanced Imaging Biomarkers.

Introduction: ADPKD is the third most common disease-causing variant in autosomal dominant polycystic kidney disease (ADPKD) after ADPKD- and ADPKD-. This study aimed to characterize the clinical presentation, progression, and distinctive imaging phenotype of ADPKD-

Methods: This retrospective cohort study included patients with disease-causing variants in , nontruncating (), or . Patients were matched by sex (48.1% male), age (mean [SD]: 57.7 ± 13.3 years), and height-adjusted total kidney volume (TKV; htTKV) (median [Q1-Q3]: 572.9 [314.1-1137.9] ml/m). Two predictive models were developed in the development cohort 81): a deep-learning model incorporating cyst-parenchymal surface area (CPSA) and cystic index, and a practical model using percentage of TKV occupied by the 2 largest cysts, with cyst volumes estimated from cyst diameters using the formula . Models were validated in an internal specificity cohort ( = 569) and an external sensitivity cohort ( = 36).

Results: Patients with ADPKD- exhibited fewer (median cyst number: 42) but larger cysts (average cyst volume: 12.1 ml), with 88.9% having no liver cysts, compared with ADPKD- and ADPKD-. The estimated glomerular filtration rate (eGFR) of decline was slower in ADPKD- (-0.69 ml/min per 1.73 m/yr) than in ADPKD- (-1.62, = 0.006) and in ADPKD- (-0.90, = 0.737). The deep-learning model demonstrated an area-under-the-curve (AUC) of 0.949 for distinguishing ADPKD- patients in the development cohort, and 88.9% specificity in the internal cohort. A volume-to-TKV ratio ≥ 18.6% identified ADPKD- with an AUC of 0.814 and demonstrated 72.2% sensitivity in the external cohort.

Conclusion: We provide a detailed characterization of the ADPKD phenotype that can be distinguished using a practical or deep-learning segmentation model applicable in diverse clinical settings.