Breast cancer accounts for 24.5% of all cancer cases in women and 2.3 million new cases are identified worldwide in 2020 (1). Among which, 18.3% of all breast cancer cases (about 420 thousand) occur in China. In addition, with advances in early detection and new therapies of breast cancer, 5-year survival has increased to near 90% and breast cancer is now treated as a chronic disease (1, 2). However, breast cancer survivors still suffer from many psychosocial burdens including stigma, depression, anxiety, fear of cancer recurrence, etc., which will result in reduced Quality of Life (QoL) (3, 4). Different from physical symptoms (i.e., lymphedema), psychosocial burdens cannot be measured by objective indicators and Patient Reported Outcomes (PROs) are developed to improve the detection of the patients’ subjective experience (5). For example, Self-Rating Depression Scale (SDS), the Center for Epidemiologic Studies Depression Scale (CES-D) and the Beck Depression Inventory (BDI) have been well developed to assess the severity of the depressive disorder (6–8). However, many existing PROs instruments are developed based on classical test theory (CTT) and focus on the accurate estimation of the symptom severity (i.e., depression, fatigue, etc.), which cannot provide detailed information at item-level. For example, most depression inventories are unidimensional and total scores are calculated to estimate the severity of the depressive symptoms with several cutoffs as reference. This procedure is straightforward but item-level information (specific symptom) is neglected, which is not consistent with definition in 10th revision of the International Classification of Diseases (ICD-10) and 5th edition of the Diagnostic and Statistical Manual of Mental Disorders (DSM-5) (9, 10).

Cognitive Diagnosis Models (CDMs), or diagnostic classification models (DCMs), are state of art psychometric models, which have been developed to classify participants into latent classes with unique profiles of attributes (11). Compared with CTT or item response theory (IRT), the CDMs provide an alternative psychometric framework for test development and score reporting. Due to the complex statistics in CDMs, most CDMs-related articles are published in the field of education measurement. However, multidiscipline researchers have been recently aware of their usefulness in PROs assessments for symptom profiles at item-level, which will be helpful to screening or intervention (12–14). Further, compared with factor analysis in CCT, CDMs allow latent attributes/symptom criteria to interact resulting in a more flexible psychometric model. Therefore, from a different perspective, the current study was designed to provide more information for the screening and monitoring of resilience as an example under the framework of CDMs. Based on the data from our previous Be Resilient to Breast Cancer (15–18), the current study was designed to compare the prediction ability of resilience to 6-month Quality of Life (QoL), between total score based on CTT and cognitive diagnostic probabilities (CDPs) based on CDMs. We hypothesized that: (1) CDMs could well extract diagnostic details about patients’ resilience strengths and weaknesses, (2) CDPs of resilience could offer incremental predictive value above conventional total score.

This is the first study to measure resilience by CDMs and its prediction of 6-month Quality of Life (QoL) in breast cancer, though CDMs have been performed to analyze PROs in pathological gambler, schizophrenia, internet addition, etc (12, 13, 34). The findings demonstrated in this manuscript confirmed the usefulness of CDMs as a sophisticated method for jointly performing resilience assessment and investigating theory-based structural facets of resilience traits. With the additional item-level information from G-DINA model, the prediction ability of resilience to 6-month QoL was significantly enhanced. Thus, the underlying structure of resilience traits or other psychological indicators (i.e., depression, anxiety, etc.) should also be considered in future PROs research. Furthermore, G-DINA-based RS-SC-10 could reduce the burden of psychologist/psychiatrist when large patients are screened or monitored.

However, two important issues in the application of CDMs should be discussed here. First, a theoretical framework or Q matrix should be set for CDMs before the formal analysis and the results will have limited diagnostic information if the theoretical framework is not relevant to the diagnosis. For example, based on our previous study, two theory-based Q matrix were developed for validating the number of item attributes and detecting the misidentified elements, including a non-bifactor Q-matrix (Model 1, Figure 2A) and a bifactor one (Model 2, Figure 2B). If both Q matrix are biased, misguided Q-matrices will lead to meaningless results with inflated slip and guess parameters, high RMSE and low Monte Carlo p-values. However, there exists no golden standard/definition for resilience and whether resilience should be conceptualized as a trait or a state is also debated (35–37). Therefore, we should be cautious about the exploratory Q matrix-based findings derived from the current study which should be replicated and validated in future research. Also, Q-matrix discovery techniques have been developed to handle with the issue about unknown Q-matrix. For example, a Q-matrix discovery method has been developed to allows for uncertainty in instrument development process by using experts’ comments in a Bayesian algorithm (38).

Second, the cognitive diagnostic probabilities (CDPs) should be validated by external indicators. In the current study, CDPs of resilience were anchored against 6-month QoL based on a prospective cohort design. However, researchers might prefer that external validation come not only from PRO instruments but also from diagnoses elicited from trained physicians. More objective indicators, i.e., cortisol, C-reactive protein, systolic pressure, etc., can be incorporated into the CDMs research, which will give more confidence to the interpretation of results (39, 40). In addition, we should be noted that CDMs take complex interactions among latent variables into consideration which allows greater flexibility than most IRT models in modeling item responses. However, this also results in complex model with too many parameters. In the current study, several reduced models have been recommended for replacing G-DINA (Figure 5) and these models can be tried in the future research to obtain more stable parameter estimates. At last, if a long item instrument with many attributes is administered in a large sample, computerized adaptive testing (CAT) can be considered to reduce the test time without a loss of measurement precision. Some research on combining CDMs and CAT has been explored in the field of psychometrics and CDMs-CAT is promising in the field of medicine (41, 42).

There are several limitations should be considered. First, many parameters are involved in G-DINA and the sample size in the current study may not be large enough resulting in potential biased estimation of Wald test for model selection or DIF detection (43). Larger sample should be considered in future research to get a robust estimated covariance matrix. Second, our sample was collected from an extremely narrow Chinese population with breast cancer and the generalizability of these findings should be further validated in other populations (i.e., lung cancer, colon-rectal cancer, etc.). In addition, CDMs allow for measurement of differences in attribute patterns cross-culturally which can help explain how cultural differences are manifested in the patterns of attributes mastered (44, 45). Third, RS-SC-10 has been validated in family caregivers (i.e., parents of children with cancer) as well as patients with different cancer types, and CDMs might also have great potentials in these vulnerable populations (46–50). More CDM-based evaluation and intervention for these vulnerable groups are urgently warranted. At last, although various CDMs have been developed to detect mastery and non-mastery of multiple fine-grained skills or attributes (i.e., DINO, DINA, ACDM, RRUM, etc.), most researchers in the field of medicine are still not familiar with CDMs due to their novelty as well as complex statistics. There exist several software programs available to estimate CDMs, for example, several R packages are available for the CDMs analyses. However, syntax preparation of these programs requires researchers’ substantial effort and more easy-use software should be developed to facilitate the application of CDMs.

Cognitive diagnostic probabilities of resilience contribute to a more accurate prediction of 6-month QoL above conventional total score. CDMs could help optimize Patient Reported Outcomes (PROs) measurement in breast cancer.

The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.

The studies involving human participants were reviewed and approved by Guangzhou University of Chinese Medicine. The patients/participants provided their written informed consent to participate in this study.

ML and PC: investigation, formal analysis, supervision, and writing—original draft. MK and AM: methodology and writing—review and editing. YT: investigation and visualization. GH and ZS: investigation. YY: data curation and software. ZY: conceptualization, funding acquisition, project administration, resources, and writing—review and editing. All authors contributed to the article and approved the submitted version.