Frequently recurrent malignant pleural effusion (MPE) significantly hampers the life quality of advanced non-small cell lung cancer (NSCLC) patients. We aimed to explore the effects of progression patterns and local intervention on MPE recurrence and apply fluorodeoxyglucose positron emission tomography/computed tomography (18F-FDG PET/CT) to establish a predictive model for MPE recurrence in NSCLC. We retrospectively recruited two cohorts of patients including treatment-naïve NSCLC diagnosed with MPE at the onset and receiving PET/CT scanning, as well as those with MPE and undergoing first-line epidermal growth factor receptor (EGFR)-tyrosine kinase inhibitor (TKI) treatment. Pleural maximum standardized uptake value (SUVmax), metabolic tumor burden (MTV), total lesion glycolysis (TLG), and uptake patterns as well as SUVmax of lymph nodes (LN) were extracted. The primary outcome was MPE recurrence defined as re-accumulation of cytologically proven ipsilateral MPE. Step-wise multivariate Cox regression was used to identify candidate variables. Cox regression analysis and random survival forest were applied to establish models. A total of 148 treatment-naïve patients with EGFR-TKI treatment and MPE were recruited during the median follow-up period of 683 days, with 69 (46.6%) and 35 (23.6%) witnessing MPE recurrence at least once and twice. Intrapleural perfusion therapy at first recurrence was a protective factor for the second MPE recurrence (P=0.006), while intrapleural perfusion therapy at baseline could not benefit the first MPE recurrence (P=0.14). Conversely, prior systemic progression indicative of the change of systemic treatment was a protective factor for time to the first MPE recurrence (P<0.001); instead, the change of systemic treatment at the first MPE recurrence was not associated with second MPE recurrence (P=0.53). In another cohort with treatment-naïve NSCLC patients with MPE and PET/CT scanning, 103 patients regardless of the actionable mutation status were recruited during the median follow-up period of 304 days. Multivariate analysis suggested that the LN SUVmax >4.50 g/mL [hazard ratio (HR), 2.54; P=0.01], female gender (HR, 0.40; P=0.01), bone metastases (HR, 3.16; P=0.001), and systemic treatment (targeted therapy vs. chemotherapy: HR, 0.32; P=0.002; immunotherapy therapy vs. chemotherapy: HR, 0.99; P=0.97) could collectively indicate MPE recurrence with an optimal 300-day area under the curve (AUC) of 0.83. For patients with actionable mutation, LN SUVmax >4.50 g/mL (P=0.02) could forecast MPE recurrence independently. In summary, LN rather than pleural metabolic activity or uptake patterns could predict MPE recurrence for patients with or without targeted therapy. We should re-consider the application of intrapleural perfusion treatment for first-onset MPE and prompt it more at the moment of recurrent MPE. Promisingly, we could probably apply the non-invasive tool to identify the risk factors for MPE recurrence.