The Radiation Oncology Feed For radiation oncologists · physicists · RTTs · dosimetrists
Worth Attention nasopharynx AI

Multimodal AI proposed major surveillance reduction after stage II NPC

External validation achieved AUC 0.986, while modeled risk-adapted surveillance reduced visits by 90.0–98.5% and identified every observed failure.

RadOncToday · ·2 min read

KEY POINTS

  • This multicenter retrospective study included 2,148 patients with stage II non-keratinizing nasopharyngeal carcinoma from five centers, treated between 2009 and 2022. Patients received IMRT alone or concurrent chemoradiotherapy, with IMRT prescribed to 68–70 Gy in 2.0–2.2-Gy fractions.
  • The study had two linked components. First, a target trial emulation used stabilized inverse-probability weighting to compare IMRT alone with concurrent chemoradiotherapy after adjustment for age, sex, EBV DNA, T stage and N stage. No significant differences in overall or failure-free survival were identified across the development and validation cohorts.
  • The second component trained a multimodal Transformer using pretreatment MRI, planning CT, three-dimensional RT dose maps, DVH parameters and clinical variables to estimate an individual patient's timing of locoregional relapse, distant metastasis or death. No post-treatment imaging or follow-up data were provided to the model at prediction.
  • Model development used 931 training patients and 192 validation patients. Independent evaluation used 658 internal-test patients and 367 patients from four external hospitals, providing a genuinely multi-institutional assessment of generalizability.
  • Internally, the model identified 45/45 failures, with sensitivity 100%, specificity 99.8%, C-index 0.938 and AUC 0.991. In the external cohort it identified 22/22 failures, while correctly classifying 342/345 failure-free patients, corresponding to sensitivity 100%, specificity 99.1%, C-index 0.851 and AUC 0.986.
  • The proposed surveillance algorithm scheduled up to five visits over five years internally and six externally around predicted failure periods. In the internal cohort, this produced 230 modeled visits versus 9,212 under the RTOG schedule—a 97.5% reduction. Against NCCN schedules, the modeled reduction ranged from 96.5% to 98.7%.
  • In the external cohort, the strategy required 150 modeled visits versus 5,138 using the RTOG schedule and 3,670–9,909 using NCCN schedules, corresponding to approximately 90.0–98.5% fewer visits. Crucially, the model assigned zero routine visits to 342 of 367 external patients who remained failure-free—an efficiency gain that is also the major reason prospective safety validation is essential.

CLINICAL TAKEAWAY

The model challenges one of oncology's most entrenched assumptions: that every survivor requires essentially the same fixed follow-up schedule. Its discrimination is striking, including external validation, but the dramatic reduction in visits exists only in retrospective simulation; no study has shown that implementing this schedule preserves timely salvage, patient reassurance or survival. Prospective testing is mandatory before surveillance can be reduced anywhere near this extent.

SOURCE

International Journal of Radiation Oncology, Biology, Physics