The LCA is particularly useful when dealing with categorical patient data because it can reveal distinct groups of patients (latent classes) that have similar treatment experiences with a comparable disease impact on their daily life.

Each class is identified based on a combination of characteristics with conditional probabilities, showing how likely it is for a variable to take certain values. The latent class must be solved for statistically independent variables to ensure that the identified subgroups are characterized only by the patterns of association.

The latent class model form is shown in Equation 1 (LCA model form):(1)pi1,i2,…,iN≈∑tTpt∏nNpin,tnwhere T denotes the number of latent classes, p_t is the unconditional probability (total sum to 1), and pin,tn is the conditional probability.

IRT is a statistical framework that can provide deeper insights into the patient-reported outcomes. The models in this family can indicate how individual items, in this case patient responses to survey questions, are associated with a latent trait.

First, LCA is used to identify subgroups of patients who respond similarly to survey items. This allows the segmentation of the patient population into meaningful clusters based on their response patterns. Second, the IRT model is fitted for the whole dataset and difficulty and discrimination parameters are estimated. Subsequently, the distribution of the latent traits is examined across the identified latent classes and then the overall item parameters are analyzed for significant variances in the observed response patterns of different latent classes.

The results of the LCA for latent class 1 are presented in Table 4. For each row, heatmapping is used to illustrate the range of conditional probability values (darker colors indicate a higher probability). This class encompasses a moderate share of the patient population (18.0%), of which a predominant group is from Germany (73.2%).

The results of the LCA for latent class 2 are presented in Table 5. For each row, heatmapping is used to illustrate the range of conditional probability values (darker colors indicate a higher probability). This class encompasses a moderate share of the patient population (18.6%), predominantly from the USA (66.7%).

The results of the LCA for latent class 3 are presented in Table 6. For each row, heatmapping is used to illustrate the range of conditional probability values (darker colors indicate a higher probability). This class encompasses a major share of the patient population (63.5%), balanced between Germany and the USA.

Patient-centric suggestions based on the observed patient feasibility latent traits:

This latent class represents mainly German patients who can benefit from improved organizational treatment aspects such as more optimized visit schedules, a reduced number of appointments per month, and the provision of some form of travel assistance.

Individuals in this class seek to join clinical trials to improve their chances for survival, but their physicians seem to have a very limited awareness of such options. An improved clinical trial design targeting this group would need a robust outreach program, an information campaign, and support from patient organizations to improve awareness. Providing additional educational materials and counseling to patients with disease progression can improve their emotional experiences and reduce anxiety.

The observations presented in this class are consistent with previous studies (19) that highlighted the importance and impact of palliative care and especially the efficacy of care delivery in glioblastoma patients with progression. Furthermore, Preusser et al. focused on the impact that a tailored strategy for symptom and complication management can have on the perceived quality of life of these patients (20).

Patient-centric suggestions based on the observed patient feasibility latent traits:

Clinical trial design for patients in the early diagnostics phase should focus on creating more meaningful and helpful interactions between healthcare providers and patients. This can include extended consultation times and more detailed treatment discussions. Similar to the previous latent class, there is very limited awareness among these patients about the opportunities for joining a clinical trial, which can be improved through pre-trial advertising campaigns and support from advocacy groups.

Although there have been few studies exploring the perspectives of newly diagnosed glioblastoma patients, the observations pertaining to latent class 2 are consistent with the findings of authors such as Fritz et al. (21), who pinpointed the role of care planning, especially in the newly diagnosed patient setting. Furthermore, our findings in this class are in concert with the role of speed in starting glioblastoma treatment and the underlying patient motivation of improving survival chances underlined by Sun et al. (22).

Patient-centric suggestions based on observed patient feasibility latent traits:

This class represents the majority of patients under active therapy. Most of them experience side effects and report dissatisfaction with treatment-related aspects, including a slower recovery, pain or discomfort, and a deterioration in sleep quality. Clinical trial design can be improved to address these side effects by providing robust patient assistance, personalized support services, and easier access to specialists. Implementing a monitoring system to identify further side effects and disease progression early can improve the patient experience and minimize the burden on mental health. Although these group members have a higher probability of joining a clinical trial than the others, the clinical trial design still needs to increase awareness significantly.

The logistical challenges can be improved through more flexible visit schedules to minimize discomfort, especially for patients experiencing side effects. Implementing technologies for remote monitoring and video conferencing might also reduce the need for patients to travel for certain evaluations.

Contrary to the initial expectation, “motivation” follows the same probability distribution across all latent classes, suggesting that it has limited informative power on improving patient feasibility compared with the other observed variables. To explore what the most informative variables are and how they discriminate and relate to individual patients, the study employs the NRM from the IRT framework.

The findings and interpretations of latent class 3 show a group of people with high scores throughout all three latent traits, indicating a developed experience within the glioblastoma treatment. Patients in this group are generally dissatisfied with the treatment objectives and are actively seeking alternative options. These findings are consistent with previous studies highlighting patterns and disparities of care in glioblastoma, especially the influence of patient motivation (23). Furthermore, pertaining to the high score of healthcare engagement present in this class, Musella et al. (24) further defined the crucial role of shared decision-making for glioblastoma patients.

Despite the lack of discrimination, two therapy types (2 and 4) are statistically significant, meaning the item is still relevant for assessing the specific characteristics that affect disease impact for patients under different treatments. In practical terms, this signifies that the question can be worded better or the categorization should be revised. •

Side effects and disease progression: The discrimination parameter for both questions is relatively high and statistically significant, implying that these items effectively distinguish between individuals based on the level of disease impact on their life. The difficulty parameters lack statistical significance, suggesting that patients easily discern whether the presence of side effects or disease progression impacts their daily life.

Examining disease impact as a latent trait shows that the questions regarding the presence of side effects, disease progression, the availability of treatment options for patients with side effects, and disease progression, as well as the reported dissatisfaction with various treatment aspects, can provide valuable insights into how to improve patient feasibility.

The item “received therapy type” also indicates a relevance to the level of disease impact; however, the question format or the data categorization must be revisited. On the other hand, the patient's country of origin and the financial impact do not seem to capture the important aspects associated with the perceived level of the latent trait (Figure 5).

The data seem to support the notion that longer patient–physician interactions lead to higher healthcare engagement. Spending more time with their physicians to discuss treatment allow patients to gain a better understanding of their treatment options and risks, allowing for more informed decision-making. It also builds trust, improving the adherence to prescribed treatment plans (Figure 6).

A patient's motivation to join a trial does not seem to affect their levels of engagement. However, data suggest that the opportunity to join a trial is of significant importance. Access to trials can contribute to higher levels of engagement because patients seek to improve their survival chances with novel treatment options and benefit from contact with highly experienced healthcare professionals, regardless of their motivation (Figure 6).

LCA and interval response theory are two statistical methods commonly used in patient-reported outcomes and patient surveys. LCA is a valuable tool for identifying distinct subgroups within a sample, which can aid in risk stratification and treatment response prediction (27).

Studies have successfully applied LCA to classify patients into different phenogroups based on outcomes after surgeries such as mitral valve surgery (27, 28). This approach allows for a more personalized understanding of patient responses and outcomes.

On the other hand, interval response theory, as demonstrated in previous studies (29), provides a method for capturing richer information by allowing respondents to provide interval-valued responses. This can be particularly useful in scenarios in which respondents may have difficulty providing precise single responses, as it allows for a range of values to be considered (29). By incorporating interval responses, more nuanced data can be obtained from patients.

Although not new, LCA and IRT statistical models have the power to use mixes of ordinal and nominal data, which are essential parts of data collection tools such as patient surveys. By harnessing the impact of IRT models, the importance and nuance within different patient-reported outcomes can be elaborated and the discriminating power of different data collection items (e.g., survey questions) can be evaluated.

In essence, while the IRT model provides a comprehensive overview of the item characteristics and latent traits observed, the latent classes from the LCA add a layer of contextual understanding that makes the surfaced insights actionable for a specific patient subgroup (30).

Although the previously mentioned models offer valuable advantages, it is also important to acknowledge their limitations. LCA and IRT models rely on certain assumptions, such as the unidimensionality and local independence of items, which may not always hold true in complex real-world scenarios (31).

In addition, the interpretation of latent classes and traits derived from these models requires careful consideration, as the results are based on statistical patterns and may not always align perfectly with clinical or practical significance. A mention in this context is the active treatment-moderate engagement class in this study. No side effects resulting from chemotherapy and radiotherapy were reported within the group; however, these patients most likely experienced some degree of side effects but probably did not mention them in the survey.

Furthermore, the generalizability of the identified latent classes and traits to broader patient populations should be approached with caution, as the specific characteristics of the sample under study may not be fully representative of other populations. Another limitation is the impracticality of applying statistical parameters from LCA models to individual new patients, which has been noted as a limitation in clinical situations (32). In addition, the use of IRT models may be limited by the number of cases and features used to derive clusters, potentially impacting the reproducibility and generalizability of the findings (33).

Another potential limitation of this study is the application of the IRT model to the whole dataset rather than within each latent class. In its current form, the IRT is applied to the whole dataset to streamline the modeling process due to the complexity of the patient survey responses. Although this approach provides broad insights into item characteristics and latent traits, it may overlook important subgroup-specific differences in item parameters that could lead to less precise insights. Future improvement in this regard would be the incorporation of Likert-type responses that would introduce a standardized measurement scale across all items (34). This would create a prerequisite for using more complex analysis by integrating IRT within each latent class (35).