AI platform for guiding personalized chemotherapy dosing feasible in clinical practice

An artificial intelligence (AI) platform that provides personalized dose recommendations for capecitabine-based chemotherapy regimens for patients with advanced cancer, including those receiving palliative care, proves useful in clinical practice, as shown in a study from Singapore.

Developed by researchers from the National University of Singapore, the platform called CURATE.AI creates a dynamic capecitabine dosing model for only one patient at a time by harnessing the patient’s own clinical data. The input data used to generate the recommendations consist of total capecitabine dose taken over the course of the cycle (the prescribed dose adjusted using the pharmacovigilance information) and the resulting response biomarker level change.

The system’s personalized dosing recommendation remains dynamic because the patient’s clinical data is constantly being fed into the system over time, the researchers said, emphasizing that this approach addresses the limitations of a maximum tolerated dose strategy.

Additionally, only small data is required, with five dose-response data pairs sufficient to generate initial profiles, they added.

High adaptability

To evaluate the performance of CURATE.AI, the researchers enrolled 10 patients with advanced solid tumours who were receiving treatment at the National University Hospital and Ng Teng Fong General Hospital. These patients had an ECOG performance status of 0–2 and were being treated with capecitabine (either as a monotherapy or in combination with oxaliplatin or irinotecan) not for curative intent.

Following enrolment, blood samples were collected to measure each patient’s biomarker levels (ie, CEA and CA19-9) before and after each treatment cycle. The patient engagement was planned for up to 12 months, with biannual survival follow-ups conducted for 3 years. The primary outcome measure was the percentage of patients in whom a CURATE.AI profile was successfully applied.

The results demonstrated that the platform had “high adaptability to clinically relevant situations encountered by patients,” the researchers noted.

CURATE.AI was considered in 73.5 percent (36/49) of all dosing events, with limited relevance for the initial dosing and complex cases. Physicians accepted 97.2 percent (35/36) of the recommended doses, with some patients receiving optimal doses that were approximately 20-percent lower on average. [NPJ Precis Oncol 2025;9:49]

This success was built on a model of human-AI collaboration, where physician feedback on the appropriate dosing range for each dosing event was incorporated into CURATE.AI operations, the researchers said. “Additionally, the dataset at the foundation of each CURATE.AI dose recommendation was co-created with the physician to ensure the relevance of the data to the specific circumstances of the patient and build trust in the CURATE.AI dose recommendations.”

However, only three patients responded to CURATE.AI-guided treatment. These patients had highly varied medical journeys, leading to different CURATE.AI profiles and their treating physicians receiving dynamically adjusted CURATE.AI dose recommendations.

As for safety, five patients had grade 3 toxicities, which were within the limitations of the small sample size and broadly comparable to known toxicity rates for the relevant chemotherapy regimens, according to the researchers. In three of these patients, the toxicities including esophagitis and non-neutropenic fever were deemed possibly related to capecitabine. Treatment was paused or delayed in these patients, with subsequent resumption. [Ann Oncol 2014;25:1356-1362]

Considering the lack of response to the selected regimens in most patients, the researchers recommended deploying CURATE.AI at a later stage, when the patient response to the regimen is confirmed/predicted (eg, through precision oncology or ex-vivo testing), in future studies.

“These are important first steps that we have made in personalizing chemotherapy drug dosing for our cancer patients. This is something that many of us as clinicians have hoped to have for our patients but has been extremely challenging to translate from idea to implementation,” said senior researcher Assoc Prof Raghav Sundar.

“The data from this research trial forms the basis for the next steps in the field of precision drug dosing in oncology,” Raghav added.