Concentration measurements
Concentration measurements of adalimumab in patient samples were performed by a validated ELISAs at Sanquin Diagnostic Services (Amsterdam, the Netherlands). In short, TNFα is captured to the ELISA plate by a coating with mouse-anti-TNFα-antibody. Hereafter adalimumab derived from the patient samples is captured and detected by a biotin labelled anti-idiotype polyclonal antibody in combination with HPR coupled to streptavidin and TMB. The concentration of anti-adalimumab antibodies (ADA) was measured by radioimmunoassay. In short, antibodies from patient samples are captured by protein A sepharose and detected with radio-labelled F(ab’)2 fragment of adalimumab. Anti-adalimumab antibodies were only measured if adalimumab concentration (back-calculated to serum) was below 5 mg/L (cascade principle). The concentrations were back-calculated to serum concentrations by taking into account the exact volume of the absorbent MitraTM tip in combination with the volume of elution buffer and a fixed haematocrit value of 0.42. For all underfilled samples, for which the correction factor was unknown, the potassium concentration was measured in eluates of completely filled tips and in eluates of the underfilled tips of the same patient. The blood volume present in the eluate was calculated with these potassium concentrations and used for calculation of adalimumab concentration.
2.6 Fit-for-purpose evaluation and statistical analysis
The predictive performance of the selected models to predict the steady state adalimumab serum concentration at 12 weeks from the measurements at day 5 and day 13 was investigated. The adalimumab concentration measured in the eluate of the MitraTM tip performed at week 12 was considered the true serum level value and compared with the individual model-predicted value. The primary outcome of this analysis was a precise and accurate prediction defined as mean prediction error (MPE) and normalised root mean square error (RMSE) < 25%. We defined normalized RMSE as RMSE divided by range (maximal dependent variable minus minimal dependent variable). Additionally, we calculated normalised RMSE defined as RMSE divided by average true values for all patients without detectable ADA. The 95% Confidence intervals (CI) are defined as 1.96 x standard error (SE) for MPE. Standard error for RMSE is defined as \(\sqrt{1/2n}\text{\ x}\) normalised RMSE, where n represents the degrees of freedom.
The clinical applicability of early prediction of steady state adalimumab levels was evaluated by dividing all predictions into four classes: true positive (prediction and measured value within therapeutic range), true negative (prediction and measured value outside therapeutic range), false positive (prediction in therapeutic range, measured value outside therapeutic range), false negative (prediction outside therapeutic range, measured value in therapeutic range).
Secondary outcome of this analysis was fitting a new model to the collected pharmacokinetic data. The pharmacokinetic parameters were estimated with NONMEM version 7.4 (ICON plc, Dublin, Ireland) and PsN version 5.2.6. (https://github.com/UUPharmacometrics/PsN) Diagnostic plots were prepared in R (R Foundation for Statistical Computing, Vienna, Austria). Model predictive ability was assessed using the proseval tool in PsN.
2.7 Ethical considerations
The study was approved by the local ethics committee and all patients provided written informed consent. The trial was registered in the Netherlands Trial Register with trial registry number NTR 7692 (www.trialregister.nl).
Results
3.1 Population pharmacokinetic model selection
Based on the literature search and the external evaluation of existing models with our retrospective dataset, the model by Ternant et alwas selected for use in this prospective analysis. Prediction corrected visual predictive check (VPC) used for the goodness of fit evaluation for this model is shown in figure 2. Other VPCs of the model as well as the model code are shown in the appendix.
3.2 Patients
A total of 56 patients were included in the trial. Drop-out rate in this trial was 20 patients (36%). Data of 36 patients were included in the prospective analysis. Inclusion was stopped at 36 patients because of the COVID pandemic. Twenty-two patients carried a diagnosis of rheumatic disease and 14 IBD. Baseline characteristics of patients included in the analysis are shown in table 1.
3.3 Fit-for-purpose evaluation
The predictive performance analysis resulted in an MPE of 294% (95% CI 261% to 326%) and a normalised RMSE of 80% (95% CI 61% to 99%). When stratified for absence of ADA, the MPE was-2.6% (95% CI -3.9% to -1.4%) and normalised RMSE 24.0% (95% CI 18.4% to 29.6%).
When calculating normalised RMSE defined as RMSE divided by average true values for patients without measured ADA, we found an RMSE of 42.5% (95% CI 37.5% to 47.6%)
Clinical applicability evaluation resulted in 75% true predictions. Full results from the clinical applicability evaluation are shown in table 2.
The results of parameter estimation based on the newly collected adalimumab levels and ADA titers collected in this study are shown in table 3.
3.3 Immunogenicity
Three patients in our cohort developed ADA at steady state 12 weeks after start of adalimumab therapy. None of these patients had received biologicals before and none of these patients were on combination therapy with other immunosuppressive drugs.
3.4 Feasibility at home
The combination of an electronic needle container and capillary blood microsampling enabled us to remotely monitor patient’s medication treatment. Exclusion from the analysis was mostly caused by home sampling errors by a minority of patients resulting in samples unsuitable for concentration measurement. Other reasons were needle drop registration issues with health beacon occurred and some patients failed to provide a complete set of three samples. These issues should be addressed to increase feasibility at home.
Discussion
In this study we demonstrated the possibility of predicting steady state adalimumab concentrations, based on early single peak and trough levels only, resulting in a correct prediction (therapeutic – subtherapeutic) in the vast majority of cases without ADA. After stratification for ADA our primary outcome measures for bias and precision were met for patients without ADA. It should be noted that ADA development is not predictable in clinical practice. The application of MAP Bayesian forecasting early in therapy in combination with an electronic needle container and home capillary sampling is unique and enables us to fully remotely monitor the patient’s medication treatment at home from pharmacokinetic point of view. Self-management can be of value for patients with chronic conditions on adalimumab treatment to reduce the number of visits to the clinic.
The foremost clinical implication of our study is the possibility of an early adalimumab dose optimisation for patients with predicted subtherapeutic levels. Since we did not measure clinical response, our prediction does not account for non-response due to other reasons.
This study shows that the population pharmacokinetic model selected (which is based on adalimumab concentrations measured in serum) could be used in combination with a VAMS method with capillary blood for adalimumab sampling. This makes sampling more accessible for patients. This method has been compared to venepuncture for adalimumab and has been studied in IBD patients at home before with reliable results.
A drawback of the current VAMS technique is underfilling of the tips. In case of underfilling, it is a challenge to calculate the concentration that equals the serum concentration. For patients with at least one correctly filled sample, other underfilled samples were corrected for volume by potassium levels in both the correctly filled sample and the underfilled sample(s). Patients with potassium-corrected samples did not perform worse in our model then patients uncorrected samples, although this could not be statistically proven due to the small number of patients. For future research with home monitoring of anti-TNF serum concentrations, a more robust sampling method (e.g. wet blood collection with microsampling tubes) is recommended to avoid these sampling and correction issues.
We used an electronic needle container to collect data on timing of adalimumab administration. Unfortunately, the electronic needle container was not able to generate a report in all cases. Therefore, on a few occasions interpolations for the timing of adalimumab administration were necessary. We do not expect this will influence the outcome of our study since adalimumab has a long terminal elimination half-life and it concerned only a single administration in a series of administrations. For implementation of our adalimumab monitoring concept, other systems such as mobile health apps may be good alternatives.
Conclusion
In this study we have demonstrated prospectively that our model is fit-for-purpose for early prediction of adalimumab levels at steady state. This concept enables early precision dosing at home to guide therapy.
Acknowledgements: We thank Wil Adriaans, Louise Merry-Meier and Antoinette Piepenbrock-van Schooten for their efforts for the inclusion of patients in this trial.
Conflict of interest disclosure: the authors declare that there is no conflict of interest
Funding information: This study was funded by Máxima Medical Center
Data availability statement: raw data were generated at Máxima Medical Center and Radboud University Medical Center. Derived data supporting the findings of this study are available from the corresponding author [PK] on request.
References
Tables
Table 1:
Patient demographics at baseline