Decoding Recurrence in Early-Stage and Locoregionally Advanced Non-Small Cell Lung Cancer: Insights From Electronic Health Records and Natural Language Processing

Abstract

Purpose: Recurrences after curative resection in early-stage and locoregionally advanced non-small cell lung cancer (NSCLC) are common, necessitating a nuanced understanding of associated risk factors. This study aimed to establish a natural language processing (NLP) system to efficiently curate recurrence data in NSCLC and analyze risk factors longitudinally.

Patients and methods: Electronic health records of 6,351 patients with NSCLC with >700,000 notes were obtained from Mount Sinai's data sets. A deep learning-based customized NLP system was developed to identify cohorts experiencing recurrence. Recurrence types and rates over time were stratified by various clinical features. Cohort description analysis, Kaplan-Meier analysis for overall recurrence-free survival (RFS) and distant metastasis-free survival (DMFS), and Cox proportional hazards analysis were performed.

Results: Of 1,295 patients with stage I-IIIA NSCLC with surgical resections, 336 patients (25.9%) experienced recurrence, as identified through NLP. The NLP system achieved a precision of 94.3%, a recall of 93%, and an F1 score of 93.5. Among 336 patients, 52.4% had local/regional recurrences, 44% distant metastases, and 3.6% unknown recurrence. RFS rates at years 1-5 were 93%, 81%, 73%, 67%, and 61%, respectively (96%, 89%, 84%, 80%, and 75% for distant metastasis). Stage-specific RFS rates at year 5 were 73% (IA), 62% (IB), 47% (IIA), 46% (IIB), and 20% (IIIA). Stage IB patients had a significantly higher likelihood of recurrence versus stage IA (adjusted hazard ratio [aHR], 1.63; P = .02). The RFS was lower in patients with clinically significant TP53 alteration (v TP53-negative or unknown significance), affecting overall RFS (aHR, 1.89; P = .007) and DMFS (aHR, 2.47; P = .009) among stage IA/IB patients.

Conclusion: Our scalable NLP system enabled us to generate real-world insights into NSCLC recurrences, paving the way for predictive models for preventing, diagnosing, and treating NSCLC recurrence.

FIG 1.

(A) Ontology of recurrence and (B) deidentified sample notes with recurrence annotations and (C) multilayer deep learning NLP architecture for recurrence curation. BiLSTM, bidirectional long short-term memory; CRF, conditional random fields; NLP, natural language processing; PET/CT, positron emission tomography/computed tomography; RLL, right lower lobe; RUL, right upper lobe.

FIG 2.

Overall recurrence-free survival and distant metastasis-free survival analysis stratified by disease (A and B) stages and (C and D) substage. Only the first recurrence events (locoregional or distant) were extracted and analyzed. Locoregional recurrence events were censored in distant metastasis-free survival analysis.

FIG 3.

Multivariable CoxPH analysis for overall recurrence-free survival (A) across all stages (I-III) and (B) subgroup of patients with disease stages IA and IB at initial diagnosis. *P < .05; **P < .01; ***P < .001. CoxPH, Cox proportional hazards; EGFR, epidermal growth factor receptor; KRAS, Kirsten rat sarcoma viral oncogene; TP53, tumor suppressor gene 53.

FIG 4.

Multivariable CoxPH analysis for distant metastasis-free survival (A) across all stages (I-III) and (B) subgroup of patients with disease stages IA and IB at initial diagnosis. *P < .05; **P < .01; ***P < .001. CoxPH, Cox proportional hazards; EGFR, epidermal growth factor receptor; KRAS, Kirsten rat sarcoma viral oncogene; TP53, tumor suppressor gene 53.