For this narrative review, we searched PubMed, Embase, Web of Science, and ClinicalTrials.gov for English-language literature published between January 2015 and October 2025. Keywords included “non-small cell lung cancer”, “perioperative”, “resectable”, “circulating tumor DNA”, “minimal residual disease”, “artificial intelligence”, and “machine learning”.

MRD generally refers to tumor cells that remain in the body after curative-intent treatment but are still below the detection limit of conventional methods^[4-7]. ctDNA is the tumor-derived fraction of circulating cell-free DNA (cfDNA) and carries molecular features of the primary tumor^[8]. It therefore provides a practical liquid-biopsy readout of residual molecular disease. In perioperative NSCLC, ctDNA-detected MRD is therefore best interpreted as a molecular surrogate of residual tumor burden after curative-intent treatment.

In resectable NSCLC, ctDNA status at the postoperative landmark time point is significantly associated with an increased risk of recurrence and death. This landmark assessment is usually performed between 2 weeks and 4 months after surgery. Gale et al. reported that patients with detectable ctDNA at this time point had a 14.8-fold higher risk of recurrence and a 5.48-fold higher risk of death^[9].

The importance of ctDNA-MRD in the perioperative setting also lies in its temporal advantage over imaging. During longitudinal follow-up, ctDNA monitoring identified 79% of patients who eventually relapsed before or at the time of radiographic progression, whereas the detection rate of a single postoperative test was only 41%^[10]. In the TRACERx study, ctDNA signaled relapse in early-stage NSCLC a median of 119 days earlier than computed tomography (CT) imaging^[11]. An exploratory analysis from ADAURA likewise showed that MRD positivity preceded imaging-defined disease-free survival events by a median of 4.7 months^[12]. Against this background, perioperative ctDNA-MRD is best understood as a risk stratification tool. It refines recurrence-risk assessment, helps distinguish biologically high-risk from low-risk postoperative states, and supports more individualized follow-up and treatment discussions.

To date, a range of MRD detection assays have been developed, and these methods differ in sensitivity, specificity, turnaround time, susceptibility to technical bias, and the degree of laboratory standardization [Table 1]^[13-18]. In practice, ctDNA-MRD detection mainly follows two broad strategies: tumor-informed and tumor-naïve, distinguished by whether prior tumor sequencing is required. Current technical approaches are mainly based on droplet digital polymerase chain reaction (ddPCR) or deep next-generation sequencing (NGS), often combined with unique molecular identifier (UMI)-based designs and other error-suppression strategies to improve interpretive reliability when ctDNA approaches the limit of detection^[19].

In resectable NSCLC, postoperative ctDNA-based MRD testing is performed in a setting in which the tumor-derived signal is often extremely scarce. After curative-intent treatment, tumor-derived DNA may fall near the lower limit of assay resolution, where conventional platforms become less stable. In the SEQC2 analytical validity study, variants above approximately 0.5% VAF were detected with high sensitivity and reproducibility across assays, whereas performance below this level became less reliable and more variable, particularly when input material was limited^[43]. This analytical constraint provides the technical rationale for AI-based ctDNA approaches. The key issue is not merely that conventional assays have limited sensitivity, but that threshold-based, locus-by-locus interpretation uses only part of the available cfDNA information. When ctDNA is sparse, weak but potentially informative signals may still be distributed across fragmentomic patterns, methylation states, and broader genomic features, even when no single locus is confidently positive^[30]. AI-based models aim to recover such information by integrating evidence across many reads and feature classes at the sample level, rather than relying solely on whether a predefined mutation exceeds a fixed cutoff^[44].

Bahado-Singh et al. provided an early example of this feature-integration strategy in a cfDNA methylation study for lung cancer detection^[45]. Using genome-wide methylation profiling with the Illumina EPIC array, they selected informative cytosine-phosphate-guanine (CpG) markers and trained several supervised learning algorithms on multivariable methylation feature sets rather than on individually thresholded loci. The cohort was small, including only 10 lung cancer cases and 20 controls, so these results require cautious interpretation. In repeated resampling analyses, several classifiers achieved area under the curves (AUCs) of 1.00, and the deep-learning model achieved 100% sensitivity and 100% specificity. These results should be interpreted cautiously, because the cohort included only 10 lung cancer cases and 20 controls, and such perfect performance in a small dataset may reflect overfitting rather than clinical maturity. Nevertheless, the study is useful as a proof of principle: when plasma cfDNA is represented as a coordinated methylation pattern, tumor-associated signal may become easier to detect than under a marker-by-marker framework^[45].

Kim et al. extended this concept by integrating methylation and fragment-size information in lung cancer plasma^[46]. Using a targeted enzymatic methylation sequencing (EM-seq) panel, they generated data from 142 patients with NSCLC and 56 healthy controls. For each sample, they constructed a two-dimensional methylation-fragment size representation and used it as input for a convolutional neural network, enabling methylation and fragment-length information to be interpreted jointly rather than as separate feature classes. The resulting model achieved an AUC of 0.87 and an accuracy of 81.5%, outperforming models based on methylation or fragment-length features alone. The dilution experiment is particularly relevant to the MRD setting: detection was achievable at a tumor fraction of 1% with 98% specificity and at 0.1% with 80% specificity. Although this framework does not by itself solve postoperative MRD detection, it supports the premise that joint modeling can extract weak cfDNA signals that may be difficult to recover when methylation, fragment size, and variant calls are analyzed separately^[46].

This direction is further illustrated by MRD-focused platforms such as MRD-EDGE. Widman et al. used plasma whole-genome sequencing (WGS) with machine-learning-guided signal enrichment to integrate single-nucleotide variant (SNV)-level and copy-number variant (CNV)-level information, rather than relying on a single feature type^[47]. In analytical testing, MRD-EDGE improved SNV signal enrichment by approximately 300-fold compared with the earlier MRDetect approach and reduced the degree of aneuploidy required for ultrasensitive CNV detection from 1 Gb to 200 Mb^[47,48]. In stage III colorectal cancer, preoperative detection reached an AUC of 0.998. More relevant to NSCLC, the same platform was applied to 22 patients with early-stage NSCLC treated with neoadjuvant durvalumab with or without stereotactic body radiation therapy (SBRT), followed by surgery. Pretreatment ctDNA detection achieved an AUC of 0.98, with ctDNA undetectable in only two patients with clinical stage IA disease. Among 14 resected patients included in the survival analysis, recurrence occurred in 5 of 8 MRD-positive patients, whereas none of the 6 MRD-negative patients relapsed during the available follow-up. Although these numbers are small, the findings suggest that whole-genome signal integration may expand the postoperative detection window in settings where conventional high-sensitivity assays remain limited^[47].

The most directly relevant evidence comes from postoperative NSCLC cohorts. Wang et al. applied WGS-based cfDNA fragmentomics to 87 patients with resected NSCLC, none of whom had received neoadjuvant therapy, and modeled fragment-size-ratio features at landmark approximately 7-day and 6-month time points^[49]. At a specificity of at least 90%, the fragmentomic models achieved sensitivities of 43.5% and 57.9% at the two respective time points, compared with 26.1% and 31.6% for a tumor-informed mutation-based assay applied to the same samples. Combining both approaches across both time points increased overall sensitivity to 78.3%, and high-risk prediction preceded radiographic recurrence by a median of 293 days. Because this study directly examined a postoperative landmark MRD setting in NSCLC, it provides one of the closest analogs to the perioperative task. It suggests that genome-wide fragment integration can recover residual-disease signals missed by locus-specific mutation tracking. However, the cohort was small, and the model was evaluated using leave-one-out cross-validation without independent external validation; therefore, these results should be considered proof-of-principle evidence rather than a definitive estimate of clinical performance^[49]. Tumor-naïve fragment-length modeling provides a complementary line of evidence. Zhu et al. developed Fragle, a deep-learning model that estimates ctDNA levels from the density distribution of cfDNA fragment lengths using low-pass WGS data^[50]. In 162 patients with resected lung cancer sampled at the landmark time point, Fragle inferred ctDNA levels without requiring tumor tissue. Among 158 patients who were classified as ctDNA-negative by a tumor-agnostic mutation panel, Fragle identified a subgroup with higher inferred ctDNA levels and significantly worse disease-free survival. These findings suggest that fragment-length patterns may provide postoperative risk information even when mutation-based plasma testing is negative. However, this analysis should be interpreted as supportive rather than definitive MRD evidence, because it represents a downstream use of a ctDNA quantification model, involved few recurrence events, and did not test CH discrimination in high-CH populations^[50].

The three principal ways in which AI/ML may enhance ctDNA-based MRD assessment in perioperative NSCLC are summarized in Figure 1.

Loss of specificity in ctDNA-based MRD does not arise solely from technical assay limitations. It also fails because plasma contains DNA from multiple biological sources, and CH is one of the most important of them. CH arises from age-associated somatic mutations in hematopoietic stem and progenitor cells, followed by clonal expansion, and many of these variants are easily detectable in plasma cfDNA. This creates a direct challenge for MRD interpretation: VAF alone often does not separate CH-derived variants from true tumor-derived ctDNA, because the two distributions overlap. Consequently, CH can produce false molecular positivity even when sequencing is analytically robust. Razavi et al. made this point clearly in pan-cancer cfDNA profiling, showing that a large fraction of plasma variants had features consistent with CH rather than tumor origin^[51]. In NSCLC, this concern has been demonstrated empirically. In the OAK cohort, Yaung et al. analyzed 332 paired plasma and PBMC samples from patients with advanced disease and found a median of one CHIP variant per sample, with counts increasing with age and TP53 emerging as the most frequently mutated gene^[34]. They also showed that allele frequencies of CHIP-derived and tumor-derived variants overlapped in plasma, which is exactly why plasma-only interpretation can become unreliable when leukocyte sequencing is not incorporated^[34]. More recent NSCLC MRD reviews have treated CH not as a side issue, but as a central obstacle to maintaining specificity when ctDNA is being used to infer residual disease after curative-intent therapy.

That is also why several groups have started to treat variant-origin assignment as a classification problem rather than a manual filtering step. Fairchild et al. pushed this further in a large real-world oncology cohort^[36]. Using labeled cfDNA variants and matched sequencing data in a training subset, they developed a machine-learning framework to distinguish CH from tumor-derived alterations and then applied it across 4,324 clinical cfDNA samples. About 30.3% of patients had evidence of CH in plasma, and the prevalence rose with age. Clinically, the key concern is not only the prevalence of CH, but also that CH-derived variants may occur in genes that would otherwise be interpreted as clinically meaningful if plasma were analyzed in isolation. The practical value of such classifiers therefore lies in reducing incorrect biological attribution at the point of reporting. In other words, the model is useful because it changes what a “positive” plasma result actually means^[36].

This logic is already evident in perioperative and longitudinal ctDNA workflows. In the LUCID study, Gale et al. used a tumor-informed design based on exome sequencing of the primary tumor to build patient-specific assays, with matched normal blood used to remove leukocyte-derived variants before plasma tracking^[9]. This should be viewed as more than a technical refinement. It is an acknowledgment that MRD interpretation becomes much less trustworthy if hematopoietic background is left unresolved from the start. Algorithmic CHIP removal combined with PBMC-based correction can likewise change which plasma calls are read as positive^[35], suggesting that CH cannot be adequately addressed by a few fixed exclusion rules alone. The more likely direction is probabilistic filtering that uses broader variant-level context and matched-normal information when available. For perioperative NSCLC, that matters because ctDNA positivity is clinically useful only if it really reflects residual tumor burden, not background hematopoietic clonality.

Before asking where AI changes this picture, it is worth fixing the clinical baseline against which any model must be judged. Gale et al. used highly sensitive, patient-specific RaDaR assays (up to 48 tumor-informed amplicons) to analyze 363 serial plasma samples from 88 patients with predominantly early-stage NSCLC^[9]. ctDNA detection at the prespecified landmark (2 weeks to 4 months after treatment) occurred in 17% of patients and was strongly associated with shorter recurrence-free survival (RFS; HR 14.8) and overall survival (OS; HR 5.48), whereas ctDNA detected 1-3 days after surgery was not associated with recurrence. That contrast matters in itself: both the timing and the persistence of detectable ctDNA, not simply its presence, carry the prognostic weight. Serial sampling showed that ctDNA-positive samples preceded radiologic recurrence by a median of ~213 days, with most detectable tumor fractions below 0.1% and over one-third of samples below 0.01%, so even a well-designed tumor-informed landmark test still operates close to its analytical floor and leaves part of the biologically relevant disease unresolved at any single time point. This is the ceiling that any downstream model has to improve on, and it frames the problem as a longitudinal one rather than a single-threshold one.

Turning these serial trajectories into an explicit, decision-grade risk score is the step where AI is genuinely relevant. Wang et al. developed PRIME, an interpretable machine-learning model trained across six curative-intent NSCLC cohorts (493 patients with stage I-III disease treated with surgery or definitive chemoradiotherapy), with MRD assessed at a postoperative landmark 2 weeks to 4 months after treatment^[52]. Rather than read ctDNA as a binary call, it combined landmark MRD with blood-based mutations in kelch-like ECH-associated protein 1 (KEAP1), serine/threonine kinase 11 (STK11) and cyclin-dependent kinase inhibitor 2A (CDKN2A) and with clinical features, after removal of germline and clonal-hematopoiesis variants. Of several algorithms, a neural network performed best, reaching an AUC of 0.85 (95%CI: 0.81-0.89) in training and 0.82 (0.74-0.89) in independent validation, with landmark MRD the dominant contributor in the SHAP analysis; model-defined high-risk patients derived a significant benefit from adjuvant therapy after surgery (P < 0.001), whereas low-risk patients did not (P = 0.928). For a perioperative reader, this is close to the intended task - a curative-intent cohort, a postoperative landmark, and an output that maps onto an adjuvant-therapy decision - but it is not a finished tool. The training population deliberately pooled resectable with unresectable and surgical with chemoradiotherapy patients, a heterogeneity the authors list first among their limitations, so its calibration in a strictly resectable, surgery-only setting remains to be shown, and the supporting evidence is still retrospective.

Zhang et al. approached the same problem from the side of the negative result^[53]. Single-time-point landmark MRD is insensitive - its false-negative rate approaches 70%, and its sensitivity is only about 30% - so LAMPAD, an XGBoost-Cox model, pairs the postoperative landmark with a preoperative baseline draw to recover prognostic information from landmark-negative patients. Trained on 163 resected stage I-III patients (no neoadjuvant therapy) with undetectable landmark MRD and validated across cohorts using both fixed-panel and personalized assays, it separated low- from high-risk patients with 2-year disease-free survival of 97.8% vs. 71.6% (HR 0.11, 95%CI: 0.06-0.21), a separation that held in pooled validation (94.3% vs. 72.4%; HR 0.18, 95%CI: 0.13-0.25), with baseline ctDNA level as the leading feature; methylation deconvolution traced the high-risk signal to immune- and multi-organ-derived cfDNA rather than tumor shedding alone. As direct perioperative evidence, this is valuable because it targets the weak negative result that most limits clinical actionability and does so in precisely the resected population of interest. The same caveats apply: the report is in press, the dual-time-point design depends on a preoperative sample that is not always banked, the validation is retrospective, and the authors note that compatibility with fragmentomic and methylation-based platforms is unproven.

The clearest demonstrations of how longitudinal ctDNA trajectories can be translated into risk scores currently come from metastatic NSCLC, where dense serial sampling is available, and AI/ML methods have been applied at scale. Assaf et al. analyzed 466 patients from the randomized phase III IMpower150 trial, sampling ctDNA at baseline and through early on-treatment timepoints, and - after algorithmic and PBMC-based removal of germline and clonal-hematopoiesis variants - trained a machine-learning model that integrated longitudinal features such as changes in VAF and the number of detectable alterations over time to predict OS^[35]. The resulting score separated patients with radiographic stable disease or partial response into high- versus low-intermediate-risk groups (hazard ratios ~3.2-3.3), with wide OS separations (for example, 7.1 months vs. 22.3 months in stable disease), and remained prognostic in an independent OAK validation cohort; in simulation, early ctDNA dynamics could even outperform early imaging for predicting trial-level OS differences. Yuan et al. reworked the same data into a treatment-agnostic, nonlinear mixed-effects joint model that links ctDNA kinetics to OS without using the treatment arm as a covariate, producing predicted survival rates generally within 20% - often within 10% - of the final trial readouts^[54]. Both are methodologically instructive, and the joint-modeling logic in particular transfers naturally to a postoperative setting. The domain does not transfer with it, however: these models were built where tumor fractions are high and serial sampling is dense, whereas perioperative MRD operates near the limit of detection and on a different endpoint, so their figures are better treated as hypothesis-generating for the resectable setting^[54].

Meta-analytic data reinforce the need to move beyond binary ctDNA endpoints when constructing such models. Lu et al. show that study-level performance improves when postoperative ctDNA is analyzed longitudinally, with AUCs approaching ~0.9 in several early-stage NSCLC subgroups^[55]. Costa et al. further argue that ctDNA VAF dynamics capture subtle changes in tumor burden that may precede both radiologic progression and clinical deterioration, making them particularly amenable to AI/ML techniques that can compress high-frequency, low-amplitude signals into continuous risk scores^[30]. Encouragingly, this is no longer a borrowed argument - the modeling is now being done in the resectable population it is meant to serve. What remains is the harder part, and it is not unique to AI: a risk score earns its place in perioperative care only once it has been tested prospectively, shown to hold across centers and assays, and tied to a decision a clinician would actually make differently. To make the evidentiary basis of this section explicit, Table 2 maps the AI/ML studies by their proximity to the target use case, separating direct perioperative NSCLC MRD evidence from findings extrapolated from screening, diagnostic, metastatic, or active-disease monitoring settings. A structured comparison of these AI/ML methods and validation features is provided in Supplementary Table 5.

In perioperative NSCLC, the clinical meaning of a post-treatment MRD result depends strongly on when the blood is drawn^[9], and any model built on these data will also be constrained by the sensitivity of the underlying assay. An early postoperative landmark sample and a later surveillance sample should not be assumed to represent the same clinical context^[32,56]. These differences suggest that indiscriminate pooling of postoperative and surveillance windows may complicate model development and interpretation. A mixed-window model may appear acceptable overall while still being less informative within specific clinical windows. A more conservative and better-supported conclusion is that predictive performance differs across postoperative windows and monitoring strategies. In localized NSCLC, landmark MRD detection yielded an NPV/PPV of 86.6%/81.0%, which increased to 96.8%/89.1% when longitudinal time points were incorporated^[32]. Existing perioperative NSCLC data support this interpretation. In resected NSCLC, postoperative sampling windows vary substantially across studies, ranging from a few days to several months, and landmark and surveillance strategies show different performance profiles, with surveillance generally improving sensitivity while changing the underlying prevalence structure of the task^[10,24]. These findings argue against treating sampling time as a purely incidental feature in perioperative MRD modeling. A more defensible approach may be to prespecify the intended clinical windows and evaluate model behavior separately within them, rather than assuming that one post-treatment setting is interchangeable with another. At a minimum, separating early postoperative assessment from later surveillance may be more clinically interpretable than applying a single cutoff across the entire perioperative course.

At the ultra-low variant allele fractions encountered in MRD analysis, the interpretation of plasma variants can be substantially confounded by CH^[51,57,58], especially in older patients, smokers, or previously treated patients^[34,59]. If plasma variants are treated indiscriminately as tumor-derived, model outputs may partly reflect CH-associated background rather than residual NSCLC alone^[60]. This distinction matters because, when CH is left unresolved, the model may no longer be answering a clean perioperative MRD question. Instead, the task may shift from “is residual NSCLC present?” to “does this patient have residual NSCLC, or a plasma background enriched for CH?” In a plasma-only pipeline, those are not the same problem. When feasible, matched white-blood-cell or PBMC sequencing remains the most robust design choice^{[34,51,58,61]}. When it cannot, CH handling should be described explicitly rather than noted only in general terms. In such cases, an explicitly described and independently evaluated in silico CH-filtering strategy is preferable to an unspecified adjustment^[62,63]. Just as importantly, perioperative AI studies would also benefit from showing how model performance changes when CH handling is omitted. Ablation analyses with and without CH filtering may be informative, because they make it possible to see whether apparent model performance is being driven by residual disease biology or by hematopoietic background. Specificity and PPV should therefore be reported transparently, especially when background non-tumor signal may influence classification. Otherwise, a model may appear accurate while assigning molecular risk on an incorrect biological basis.

Fragmentomic MRD features are influenced by both biological and technical context. Their observed values can be shaped by both biology and the analytical pipeline, although the size of this effect is feature-specific rather than universal^[64-66]. In cfDNA fragmentomics, preanalytical and analytical factors such as collection and processing conditions, library preparation, and bioinformatic processing can alter some readouts, particularly genome-wide fragmentation patterns and fragment-end features^[66]. By contrast, not every variable perturbs every feature to the same degree; for example, collection tube and processing time did not materially change overall fragment-size distributions in one study, whereas genome-wide fragmentation patterns and fragment-end sequences were affected, and one- versus two-step centrifugation had little effect on nuclear cfDNA size or motif diversity but did affect mitochondrial cfDNA^[65]. Wang et al. made this problem especially clear by comparing donor-derived plasma cfDNA processed with nine library kits and ten data-processing routes: kit- and route-dependent variation was evident, particularly for motif profiles and some regional length signals, and batch effects remained visible across published datasets even after standardized reprocessing^[66]. In perioperative NSCLC, where post-treatment ctDNA can be present at extremely low fractions, such technical shifts cannot be dismissed, because assay-dependent variation may be mistaken for weak biological signal^[33]. This does not imply that every strong single-center model is non-portable, but it does indicate that portability should not be inferred from internal performance alone^[67,68]. Cross-study fragmentomics work has specifically highlighted the need for external validation, and claims of generalizability are better supported when models are evaluated in independent site-held-out data. For reporting, the most defensible standard is to describe the assay and preprocessing pipeline in enough detail for reproduction, identify the development and validation settings, and report discrimination and calibration transparently, including any heterogeneity across clinical windows or centers^[69,70]. By contrast, pipeline locking, version control, and longitudinal drift monitoring are best framed as strong translational recommendations rather than as conclusions already established by perioperative NSCLC MRD studies.

Many published fragmentomic AI models for lung cancer have been trained in screening, diagnostic, or active-disease treatment-monitoring cohorts rather than in postoperative MRD cohorts^[71-73]. Models developed directly within resected, postoperative NSCLC populations remain few, and the existing examples are typically limited by small cohort size and an absence of independent validation^[49]. Perioperative MRD in resectable NSCLC belongs to a different domain, with post-treatment ctDNA often approaching the analytical limits of detection and the task shifting from current cancer detection to future recurrence prediction. For this reason, perioperative AI-MRD models should ideally be developed in genuine perioperative cohorts, or at minimum externally validated and, when necessary, recalibrated in such cohorts, rather than transferred unchanged from screening or active-disease settings^[74]. The target population, outcome definition, and signal-to-noise regime are different, and postoperative samples generally show a much weaker tumor signal than cancer-versus-control or active-disease settings. A model that behaves sensibly in a “cancer versus control” setting may therefore become poorly calibrated once it is moved into a postoperative relapse-prediction setting^[67,72]. In practice, this means that recalibration and threshold re-estimation may be needed after external validation, especially when models are moved across clinical windows or domains.

In perioperative NSCLC, an MRD score becomes clinically useful only when it is linked to a prespecified clinical action. A probability estimate alone is insufficient. This is one reason why strong retrospective performance does not automatically translate into practice^[31,75-77]. The postoperative MRD literature in NSCLC remains heterogeneous, there is still no single standard-of-care framework for how liquid-biopsy MRD should be deployed, and recent meta-analytic work continues to show that performance depends on assay type, monitoring strategy, and the clinical context in which the test is applied^[78-80]. Thus, the central task is not only to build a model that ranks risk, but also to define what action should follow each risk level. One practical route is to align the score with the target perioperative population and map the output into a small number of prespecified action categories: closer surveillance, continuation of standard therapy, consideration of escalation, or in selected contexts de-escalation^[81]. The goal is not to force every score into an intervention. It is to prevent the opposite problem, where the same model output may be interpreted differently across centers because no action framework was set in advance. Without that calibration-plus-action structure, even a statistically strong MRD model is likely to generate variable reporting, variable management, and very little workflow benefit^[31]. In other words, the model may predict risk, but it will not support reproducible care unless its outputs are linked to predefined clinical actions^[75].

The major sources of uncertainty and potential AI/ML failure modes across the perioperative ctDNA-MRD workflow are summarized in Figure 2.