Small cell lung cancer (SCLC) is an aggressive and challenging disease to treat, mainly due to the lack of knowledge surrounding chemoresistance mechanisms [1,2]. Standard SCLC treatment consists of defined treatment lines, each comprising 4-6 cycles of platinum-based chemotherapy (cisplatin or carboplatin) with etoposide (PE) or irinotecan (IP). For patients with cancer stages I-III, this is often combined with concurrent thoracic radiotherapy. In stage IV cancer, immunotherapy with checkpoint inhibitors can be combined with chemotherapy, though with a modest median overall survival gain of 2‐3 months in randomized controlled trials (RCTs) and with even smaller survival benefits reported in real-world studies [3,4]. However, most patients with SCLC experience relapse within 6 months after initial treatment, necessitating further therapy, though with limited survival. Real-world studies further indicate that outcomes beyond second-line therapy are poor [5], and the 5-year overall survival for patients with SCLC remains below 10% [6].

In the setting of disease relapse, the distinction between sensitive-relapse and resistance-relapse plays a crucial role in predicting survival. Patients with sensitive disease, defined by relapse occurring more than 90 days after the completion of first-line platinum-based chemotherapy, might be rechallenged with the same chemotherapy regimen administered as first-line and demonstrate significantly better survival outcomes compared to those with resistance disease, where relapse occurs within 90 days [7,8]. In this latter case, the most common approach is to change to other chemotherapy agents if the patient is still in acceptable general condition.

The rapid progression of SCLC often makes it difficult to manage [9], which results in discrepancies in observed outcomes between clinical trials and real-world patients [10]. The knowledge gap between clinical trials and the real-world underrepresented patient populations can be overcome by treatment adjustments based on clinical experience, including choice of chemotherapy agents, dose adjustments, and end-of-life care [9,11,12]. The focus of previous studies has mainly been on survival outcomes, often overlooking patients who received subsequent treatment lines after first-line chemotherapy. Further, there is a lack of studies that analyze events that occur over time in the treatment process [9,13-15].

Real-world data, here defined as health care data produced for the primary purpose of treatment in a clinical setting, offer an opportunity to retrospectively explore the impact of treatment decisions and pathways on patient outcomes, given the variability in the real-world patient population [16,17]. As such, real-world data have the potential to inform the individualization of treatments for specific subpopulations [18,19], such as patients with poor Eastern Cooperative Oncology Group performance status (ECOG PS), older patients, and patients with brain metastases who tend to be underrepresented in clinical trials [10].

However, evidence generation from real-world data is challenging because of complex clinical workflows and issues related to its secondary use, such as inherent biases in data collection and decision-making processes, and nonobservable effects [20,21]. In SCLC treatment, the nuances of dose adjustments, planned treatment pauses, assessment of response, and early discontinuation of treatment due to severe toxicity or health deterioration exemplify the intricacies involved [22,23].

The aim of this study is to describe the time dynamics of patient trajectories in a clinical patient with SCLC cohort in order to identify factors of importance for dose adjustments and survival outcomes. This approach necessitates an appropriate choice of data analysis [24]. While artificial intelligence and machine learning have shown much promise for analyzing real-world data, issues of transparency and interpretability pose a potential risk of propagating biases from the data, including not-captured information and operational confounders [10,25].

Here, the focus was on selecting transparent modeling approaches [25,26], allowing for clinical contextualization and interpretation [10,21]. Then, the analysis consisted of a combination of Bayesian [27-29] and mixed effects models [11,30,31] for studying dose adjustments and their impact on adverse events (AEs; Bayesian logistic regression [32]) and the grade of toxicity (Bayesian cumulative link ordinal regression [32]). Competitive risk Fine-Gray regression [33] was applied to compute cause-specific hazard ratios (csHR) and predict the time course of the therapy, ECOG PS deterioration, and early discontinuation of treatment, defined as discontinuation before completing 4 planned chemotherapy cycles due to AEs, or radiological and clinical signs of tumor progression.

In this study, we analyzed a longitudinal real-world cohort of patients with SCLC and their treatment pathways. The analysis allowed the study of the relationships between treatment decisions, phases of progression of disease, occurrence of AEs, and overall survival. To the best of our knowledge, this is one of the most detailed longitudinal datasets of SCLC real-world treatment decisions. This type of data aims to add novel insights on dosing effects over time, the emergence of toxicity, and the decision points relative to the clinical outcomes.

The improved prognosis of patients who experienced neutropenia aligns with previous findings [43,44]. The analysis allowed the investigation of treatment-related AEs in relation to the efficacy of therapy or to detect when this was instead linked to deterioration of the patient’s conditions. For example, the cumulative incidence curves of AEs stratified by responding patients against early discontinuation of the therapy in Figure 2 showed the role of neutropenia when the time windows of AEs may overlap with the efficacy of treatment (Figure 3).

We identified significant ORs relating to the dose of etoposide and increased occurrence of AEs. The analysis showed a similar OS for patients receiving 4 doses during first-line therapy, regardless of whether the patient had a dose adjustment/reduction or not (Figure 3). A worse prognosis was observed when there was a deterioration in ECOG PS or early discontinuation of treatment.

We interpret our findings as indicating that the occurrence of neutropenia reflects a longer duration of treatment exposure, rather than being a causal factor for improved survival. Patients who completed all 4 treatment cycles—often at reduced doses—demonstrated better outcomes than those receiving fewer cycles at higher doses. This is supported by the higher incidence of neutropenia among patients with treatment response, suggesting that extended exposure, not intensity, is the key contributor to therapeutic benefit.

SCLC treatment effects vary widely between individuals. This could suggest the need to identify key covariates and clinically relevant subgroups that might inform a more tailored dosing, or in some cases, guide decisions against treatment altogether, in order to improve survival while avoiding AEs in patients unlikely to benefit from therapy [11,12,45]. Indeed, etoposide dose calculation is based on BSA, which has been shown to present such limitations [46]. Similar analyses could be done for the second-line therapy. In fact, only a few patients received the full expected dose without dose reduction, and the platinum dose and PE treatment were strongly associated with AEs.

ECOG PS 2‐3 and elevated LDH serum levels are known to be prognostic factors of worse OS in SCLC [9,47]. Our results, obtained by the careful analysis of the longitudinal data line by line with the clinical experts, added a new insight by showing their role in predicting early discontinuation of treatment, which was associated with significantly poorer OS.

Another finding was the high probability of a favorable response to first-line therapy in male patients, along with their lower likelihood of experiencing AEs (Tables 3 and 4; Tables S4 and S5 in Multimedia Appendix 1). Moreover, when stratifying older patients by sex, women who developed neutropenia exhibited survival rates comparable to those in the same subgroup who did not receive chemotherapy. A possible explanation is that female patients have historically been underrepresented in clinical trials [10,48,49]. In addition, in the second-line setting, the worse PFS and high frequency of early discontinuation for male patients might suggest the need for further studies regarding different therapy responses for males and females.

The high-risk subgroup analysis allowed detection of key groups of patients such as older patients with poor ECOG PS, where an eventual poor response to therapy is associated with similar survival as compared to patients not receiving chemotherapy. This finding is clinically relevant, as it highlights the importance of balancing the benefit and risk of treating such fragile patients and may help clinicians avoid overtreating this vulnerable population. These outcomes underline the potential role that real-world data may fill in providing evidence on treatment in groups of patients that are underrepresented in clinical trials [10].

The design of the analytic pipeline relied on the consideration that the data should be analyzed using models capable of capturing the complexity of the patient treatment pathways without sacrificing interpretability (such as black box machine learning models [25]). We believe that this is an important consideration when approaching real-world data. In fact, the discussions with the clinicians on the obtained results allowed the assessment of the validity of the results, the detection of potential inherent bias, and confounders related to operations and nonobservable effects not captured from the data. For example, it was observed that the 10-day dosing regimen of etoposide resulted in poorer survival. At the time of the present analysis, this dosing regimen had, however, already been abandoned due to toxicity caused by the limited time for patients to recover between the cycles (Tables 2 and 3).

The study had several limitations. The reason for missing values was not at random for some of the variables, thus introducing some biases in the results. Particularly for baseline neutrophil counts, where missing observations were high across all treatment lines. This bias is partially mitigated by the MissForest algorithm because it is a nonparametric model, and the algorithm does not need an explicit assumption about the missing mechanism (ie, missing at random) [38].

Further, these variables were available only at the start of the treatment line, meaning that we were not able to analyze their time profiles (ie, neutrophil counts over time between doses). Moreover, it should be stated that AEs are not always described in medical notes, thus introducing investigator bias in the collected data.

Another limitation was the sample size of the high-risk subgroups and patients receiving the second-line treatment, resulting in higher estimates of variability of the outcomes as compared to the first-line treatment cohort. In detail, the odds ratios of the mixed effects models resulted in wide confidence intervals. The reason for this high variability in outcomes could be that these small cohorts were also associated with fewer longitudinal observations per patient (the number of administered doses). Increasing the sample sizes of these subgroups in the future by collecting more data will help to provide more precise estimates of the covariate effects. Moreover, patients receiving more than two lines were not included in the multivariate analysis, as this was only observed in a few instances.

The median age of our population was 72 years, yet we chose 75 years to better isolate a more vulnerable subset of patients by moving further from the median. This aligns with existing studies that identify this age group as particularly fragile, often requiring specialized considerations in treatment planning [50]. Future studies exploring different criteria and strategies to separate older patients would be required.

This was a single-center study, and expanding the analysis by including other SCLC cohorts from other centers would be beneficial. A comparison of the dose adaptation and the follow-up processes in clinical trials would be of high importance [10]. A future aspect to explore would be to expand the collection of longitudinal real-world data, designed similarly to the study presented here, in larger initiatives such as multicenter studies or cancer registries [51-53], and recent new proposed therapies (ie, durvalumab after chemoradiotherapy [54], tarlatamab/lubri in second- and third-line settings [55], and immune-oncology therapies [9]).

This study could inform the design of future works, such as multistate models of patient survival in function of different therapies and events [56,57], or model-informed precision dosing models to explore individualized treatment and dosing to reduce the risk of AEs (eg, neutropenia) without compromising the efficacy of the treatment in terms of survival outcomes [58,59]. Furthermore, the outcomes from the approaches presented herein can inform more reliable future analyses using, for example, machine learning algorithms suitable for modeling longitudinal processes, such as physics-based neural networks [60].

A continued fruitful collaboration between clinical and technical experts when analyzing complex data and contextualizing the clinical processes, as we did in this work, would enhance the impact of real-world data in clinical practice and clinical drug development.