1. The UCSF Health problem.

Tacrolimus is a major immunosuppressive drug in solid organ transplantation.¹ Due to its narrow therapeutic index, however, tacrolimus administration requires strict monitoring and adjustment to achieve optimal therapeutic dosages. Excessive dosages may increase the risk of nephrotoxicity while insufficient dosages can lead to acute rejection.² Consequently, transplant patients require lifelong monitoring of tacrolimus trough concentrations.

When trough levels fall outside the desired range, clinicians must account for a patient’s last dosage, including the dosage amount, the route and timing of administration, and concomitant medications. Other factors, such as gene polymorphisms, comorbidities, and patient demographics impact tacrolimus pharmacokinetics.³ As a result, dosages required to reach target whole blood concentrations of tacrolimus vary among individuals. This variability imposes substantial time burdens on practitioners who may struggle to account for all relevant covariates. Patients also face a significant cumulative time burden.⁴ Following their first admission, patients must travel to the hospital twice weekly as outpatients for venous blood sampling, which are required to review therapeutic drug monitoring results.⁵

While previous studies have explored ML and tacrolimus trough levels in kidney^6,7 and liver^8,9 transplant recipients, data on tacrolimus dosage prediction is limited. Yoon et al.⁹ developed a long short-term memory (LSTM) model in liver transplant recipients based on trough levels up to 14 days following transplantation. However, the model did not adjust for patients with once-daily dosage, concomitant medications, and comorbidities. Additionally, there are few studies that have explored machine learning approaches in tacrolimus dosing in lung¹⁰ and heart transplant recipients. Our model intends to fill this existing gap while improving patient safety and reducing costs associated with dosing errors.

2. How might AI help?

Long short-term memory (LSTM) is a recurrent neural network that processes sequential data to generate predictions, and LSTM models have been previously used to predict therapeutic tacrolimus concentrations.

By proposing accurate tacrolimus dosages, AI will save the time clinicians spend working through individual patient data in the immediate postoperative period, reduce the high-level of expertise to make clinical decisions, and additional patient lab checks, which will ultimately save clinical costs. Moreover, by preventing potential over- and underdosing, AI will help mitigate patient hospital length of stay and readmissions. Indeed, previous studies demonstrate that personalized tacrolimus dosing over time leads to shorter median hospital stays compared to conventional dosing.⁹

We propose developing a LSTM model that predicts tacrolimus concentrations and proposes dosage adjustments using data from the EHR “Patient Synopsis” and “Transplant Summaries” from APeX. Data will be extracted using Versa with assistance from the UCSF Data Core or through Transplant Insights, which can be used to scan both UCSF and external EHRs. We will begin by implementing tacrolimus dosage monitoring for kidney transplant recipients, given the large cohort size, number of touchpoints, and the quality of data available, which will be sufficient to train our models. Patients must complete daily trough-level testing during the initial 10-day hospitalization and twice-weekly outpatient testing for months. With nearly 400 annual kidney transplants, this imposes significant financial burdens and demands substantial time commitments. Our model will then be extrapolated to lung, liver, and heart transplant recipients.

Our goal is to implement a clinical decision support tool to guide providers in adjusting tacrolimus dosages in the immediate postoperative phase and during ongoing immunosuppression maintenance of solid organ transplant recipients. This would be a practice-changing technology that would save time for physicians, patients, and caregivers by (1) assisting in clinical decision making and ultimately (2) limiting patient trips to the lab. Additionally, by accounting for all covariates, LSTM models will help reduce potential human error and improve patient safety by preventing over- or underdosing.

3. How would an end-user find and use it?

The AI tool will be integrated with UCSF’s APeX EHR system in a unique tab. This tab will present predictive tacrolimus dosage with a streamlined workflow. The workflow will show trough trends, current tacrolimus medication regimens, correlation with clinical stays, and allow clinicians to quickly adjust covariates (e.g., concomitant medication modifications and discontinuations). Using the model, clinicians can specify the desired tacrolimus level for a patient, and the appropriate dose will be recommended. Additionally, clinicians will be able to pull out and examine aggregate trends for their patients.

 4. Embed a picture of what the AI tool might look like.

Figure 1 depicts what the AI tool may look like. This will be available as a separate tab in APeX. The transplanted organ will be shown. Directly below the transplant information, previous model predictions can be tracked to verify accuracy and whether they were followed. The predicted tacrolimus projected dose is shown as a star and will estimate the expected trough. As users input information in the right panel, “Adjust for Your Next Dosage”, they will have the option to alter data as necessary. Weight, height, and age will be automated, but users will have the option to enter manual data, if necessary, as indicated by the “More” button. Concomitant medications can be selected from a patient’s current medication. Labs will be automatically entered in the left panel.

Figure 1. Proposed model. Please note that hospital stays are not reported, but the model would include this feature.

5. What are the risks of AI errors?

AI errors in the context of LSTM models include overfitting or underfitting on training data. This may lead to inaccurate tacrolimus dosage recommendations that compromise patient safety by through toxicity or insufficient immunosuppression. Previous data, however, has suggested that LSTM models can accurately predict tacrolimus dosages to achieve actual concentrations within the therapeutic range when sufficient training data is provided.⁹

Despite these risks, UCSF maintains complete data on a large cohort of transplant recipients, which will be used to validate the model. The model will undergo continuous performance monitoring, bias assessment, and iterative refinements. We plan to incorporate algorithms to enhance model transparency and observe the effects of covariates on tacrolimus concentrations. Additionally, our tool will assist clinical decision making; clinicians will use their discretion when administering tacrolimus dosages.

6. How will we measure success?

We will measure the success of our model for tacrolimus dosages based on adoption, the impact on clinician efficiency, and patient outcomes. These metrics include the following:

Primary Outcome —Therapeutic Accuracy: We will track how consistently patients’ tacrolimus levels remain within the desired therapeutic range, as well as the rate of dosing adjustments that adhere to AI recommendations.

Secondary Outcomes:

Healthcare Costs: We will analyze overall costs related to tacrolimus therapy, including additional laboratory tests, ICU/hospital length of stay, pharmacist and clinician charting durations, and hospital readmissions associated with suboptimal dosing (toxicity or organ rejection). These measures will be compared pre- and post-implementation.

Clinical Efficiency: We will examine the frequency of clinic visits for dosage adjustments, readmission rates due to organ rejection or nephrotoxicity, and overall hospital length of stay will be compared pre- and post-implementation.

User Feedback and Engagement: We will solicit feedback from transplant clinicians on their experiences integrating the AI tool into daily practice. This may include quarterly Qualtrics surveys focused on usability, perceived accuracy, and impact on workflow.

Adoption Metrics: We will measure how often the AI-driven recommendations are accepted or overridden, as well as the reasons behind provider decisions.

Continuous Improvement: Feedback loops will be established to refine the AI model, ensuring that user input guides enhancements and fosters long-term adoption.

7. Describe your qualifications and commitment:

Dr. Steven Hays is a pulmonologist and director of the UCSF Lung Transplant Program and Medical Director of the Transplant Digital Health team. The team includes Anna Mello, the Manager of Transplant Quality and Digital Health; an acute care nurse practitioner; transplant coordinators; and Logan Pierce. Dr. Pierce serves as Managing Director for the Department of Medicine Data Core, which focuses on clinical data extraction, analysis, and visualization.

The Transplant Digital Health team has launched several projects, including the UCSF Health Home Spirometry Kit, a self-sustaining project that generates $1.2 million per year. Additional projects include Transplant Insights, a tool that pulls information from 95% of EMRs to create clinical summaries; and The Kidney Pre-List Chat Program, which enabled data-driven triage of online referrals and an annual savings of upwards of $580,000.

Additional team members include Dr. John Roberts, a board-certified surgeon who specializes in abdominal transplantation, and Dr. David Quan, a transplant pharmacist. Dr. Roberts has published nearly 170 papers on topics such as immunosuppression. Dr. Quan oversees UCSF's transplant pharmacist group and serves as program director for the UCSF Medical Center's specialized residency in solid organ transplant.

References

1.         Araya AA, Tasnif Y. Tacrolimus. In: StatPearls. StatPearls Publishing; 2025. Accessed April 3, 2025. http://www.ncbi.nlm.nih.gov/books/NBK544318/

2.         Randhawa PS, Starzl TE, Demetris AJ. Tacrolimus (FK506)-Associated Renal Pathology. Adv Anat Pathol. 1997;4(4):265-276.

3.         Staatz CE, Tett SE. Clinical Pharmacokinetics and Pharmacodynamics of Tacrolimus in Solid Organ Transplantation. Clin Pharmacokinet. 2004;43(10):623-653. doi:10.2165/00003088-200443100-00001

4.         Veenhof H, van Boven JFM, van der Voort A, Berger SP, Bakker SJL, Touw DJ. Effects, costs and implementation of monitoring kidney transplant patients’ tacrolimus levels with dried blood spot sampling: A randomized controlled hybrid implementation trial. Br J Clin Pharmacol. 2020;86(7):1357-1366. doi:10.1111/bcp.14249

5.         Leard LE, Blebea C. The transformation of transplant medicine with artificial intelligence-assisted tacrolimus dosing. J Heart Lung Transplant. 2025;44(3):362-363. doi:10.1016/j.healun.2024.11.029

6.         Mok S, Park SC, Yun SS, Park YJ, Sin D, Hyun JK. Optimizing Tacrolimus Dosing During Hospitalization After Kidney Transplantation: A Comparative Model Analysis. Ann Transplant. 2025;30:e947768. doi:10.12659/AOT.947768

7.         Zhang Q, Tian X, Chen G, et al. A Prediction Model for Tacrolimus Daily Dose in Kidney Transplant Recipients With Machine Learning and Deep Learning Techniques. Front Med. 2022;9:813117. doi:10.3389/fmed.2022.813117

8.         Li ZR, Li RD, Niu WJ, et al. Population Pharmacokinetic Modeling Combined With Machine Learning Approach Improved Tacrolimus Trough Concentration Prediction in Chinese Adult Liver Transplant Recipients. J Clin Pharmacol. 2023;63(3):314-325. doi:10.1002/jcph.2156

9.         Yoon SB, Lee JM, Jung CW, et al. Machine-learning model to predict the tacrolimus concentration and suggest optimal dose in liver transplantation recipients: a multicenter retrospective cohort study. Sci Rep. 2024;14(1):19996. doi:10.1038/s41598-024-71032-y

10.       Choshi H, Miyoshi K, Tanioka M, et al. Long short-term memory algorithm for personalized tacrolimus dosing: A simple and effective time series forecasting approach post-lung transplantation. J Heart Lung Transplant. 2025;44(3):351-361. doi:10.1016/j.healun.2024.10.026