An analysis of patient-reported outcomes (PRO) in breast cancer patients through item-response theory (IRT) pharmacometric modeling

Aims: Patient-reported outcomes (PRO) provide a valuable means to assess the subjective impact of a disease and treatment on patients’ symptoms, adverse effects, functional status and quality of life. An item response theory (IRT) pharmacometric (PM) approach (1) is proposed herein to facilitate PRO interpretation, accounting for their multi-scale nature and frequent missing data. This IRT PM analysis aims to characterize Functional Assessment of Cancer Therapy-Breast (FACT-B) data in breast cancer patients following treatment with trastuzumab emtansine (T-DM1), investigate potential exposure-response relationships, and compare the response of TDM-1 to a reference (capecitabine and lapatinib) treatment.

Methods: The FACT-B questionnaire consists of 36 items with ordered categorical answers, divided into 5 subscales: physical, social, emotional and functional well-being, and a breast-cancer subscale (BCS). Item-level FACT-B data were collected at multiple visits from locally-advanced or metastatic breast cancer patients involved in a phase 3 trial (2) and treated with TDM-1 (N=484) or capecitabine and lapatinib (reference arm, N=478). The IRT longitudinal model was developed in three steps using TDM-1 arm data. In step 1, a base IRT model was fitted to naively pooled data from all visits. Proportional-odds models described the probability of each item’s score as a function of item-specific parameters and a latent variable (well-being, Ψ) specific to each patient, visit and subscale. Ψ was assumed to be standard normally distributed at baseline. The empirical Bayes estimates of Ψ were subsequently used as dependent variable in step 2 where a longitudinal well-being model was developed to describe Ψ individual time-courses. Linear and non-linear functions of time were considered. A stepwise covariate search evaluated the effect of baseline characteristics (demographics, disease state and prior therapies) on model parameters. TDM-1 exposure (AUCcycle1 and Cmin,cycle1)(3) effect on model parameters was also investigated. In step 3, models developed in steps 1 and 2 were combined into a final longitudinal IRT model. A similar 3-step approach was applied to the reference arm data, fixing the item-specific parameters and keeping the longitudinal model and covariate model structures as in TDM-1 model.

Results: In the base IRT model, 180 item-specific parameters were estimated using TDM-1 data. The BCS being heterogeneous, its items may thus not relate to the same underlying variable and could not be adequately fitted using a separate Ψ variable; their reassignment to one of the four other subscales based on log-likelihood ratio tests resulted in an improved fit (dOFV=-1138). Correlations at the individual level between Ψ from different subscales ranged 34-68%. The time course of Ψs was best described by an asymptotic function of time. A large additive inter-individual and moderate inter-subscale variability on the asymptote parameter Ψss allowed patients to improve (Ψss>0), stay stable (Ψss=0) or worsen (Ψss<0) in a subscale-specific manner. A common progression half-life for all subscales was estimated to 52 days. The typical emotional well-being improved over time whereas social well-being worsened. No statistically significant typical change in well-being was identified for the physical and functional subscales. A statistically significant (p>0.001) effect of race was found on the baseline social and functional well-being. No relation was identified between progression and TDM-1 exposure. When applied to reference arm data, the IRT longitudinal model showed a statistically significant typical worsening in the physical subscale resulting in lower typical expected scores for the physical subscale items in the reference arm compared to the TDM-1 arm. The typical asymptote was also slightly lower for the other subscales.

Conclusion: The developed framework adequately described longitudinal FACT-B data following TDM-1 or reference treatment. It could characterize the multidimensional nature of FACT-B and handled missing scores with no need for imputation. No exposure-response relationship was found in the TDM-1 arm for either of the subscales. There was a worsening of physical well-being identified in the chemotherapy containing reference arm, whereas in the TDM-1 arm patients typically stayed stable. When combined with traditional efficacy and safety analyses, PRO offer health care professionals, patients and payers an additional assessment of overall health care value from a patient-centered perspective at the approved dose.


1. Ueckert et al. Pharm Res. 2014;31(8):2152-65.

2. Welslau et al. Cancer. 2014;120(5):642-51.

3. Lu et al. Cancer Chemother Pharmacol. 2014;74(2):399–410.