Drug resistance is the main obstacle to achieving a cure in patients with cancer (1). Multiple mechanisms are associated with developing resistance, which presents a major therapeutic challenge (2). The current methods for overcoming resistance are far from satisfactory (3).

Thyroid cancer is the fastest-increasing malignancy in the United States and the eighth most common cancer globally (4). The treatment of its most common form, differentiated thyroid cancer (DTC), consists of surgery, radiotherapy, and radioactive iodine (RAI) (5). Up to 15% of DTCs are refractory to RAI and carry a poor prognosis (5, 6). Lenvatinib, a multi-targeted tyrosine kinase inhibitor, was approved as a monotherapy for treating locally advanced or metastatic RAI refractory DTC (RR-DTC) (5). Lenvatinib is also approved for advanced renal cell carcinoma, endometrial carcinoma, and hepatocellular carcinoma (7). Despite its efficacy, lenvatinib is associated with the development of drug resistance and with a spectrum of adverse events (AEs), including hypertension, fatigue, proteinuria, and gastrointestinal disturbances, which often necessitate dose reduction, interruption, or permanent discontinuation (5, 6, 8–10). Traditional strategies to overcome resistance, such as combination therapies and dose escalation, have not been widely implemented into clinical practice for DTC treatment and may be accompanied by increased toxicity (10). Further, a temporary cessation of treatment (i.e., a “drug holiday”) carries the risk of cancer progression.

The Constrained Disorder Principle (CDP) defines complex systems by their degree of inherent variability (11). Integrating CDP-based second-generation artificial intelligence (AI) systems into treatment regimens has shown promising results in enhancing therapeutic outcomes, overcoming drug resistance, and minimizing AEs (12–17). Personalized variability-based regimens optimized drug effectiveness in patients with congestive heart failure, multiple sclerosis, and chronic pain (13–15, 17–20).

The present study aimed to assess for the first time the effect of personalized dynamic adjustment of lenvatinib dosages and administration timing, guided by an AI-driven mobile application, on drug efficacy and AEs in patients with drug resistance or intolerance.

The data of this proof-of-concept feasibility clinical trial showed that it is possible to overcome resistance to lenvatinib using a CDP-based second-generation AI system. The data supports the safety of the system and its ability to improve clinical outcomes. The disease control rate, comprising complete response, partial response, and stable disease cases, was 80% within a 14-week follow-up. The FDG-PET scan-based RECIST v1.1 and PERCIST criteria showed partial response in 40% of patients ( Figure 1 ) and stable disease in an additional 40% of the patients. One patient experienced a progressing disease. 75% of the patients with thyroid cancer showed a reduction in thyroglobulin levels, and 60% of patients showed a decrease in NLR during treatment ( Figure 2 ). Improvement in the median social support score among patients utilizing the system supports an ancillary benefit of the intervention. No grade 4 AEs were recorded, no ER visits occurred, and no functional deterioration was noted.

Resistance to lenvatinib is a significant challenge that may cause loss of response to treatment through various molecular mechanisms related to cell death regulation, metabolism, histological transformation, and epigenetics (9, 10). Currently, methods are lacking to differentiate actual resistance from loss of response due to inadequate dosing, such as serum drug level monitoring. The digital pill, regulated by a CDP-based second-generation AI system, introduces an innovative approach to drug administration by incorporating variability signatures (15). Biological systems inherently display variability, which is mandatory for their proper function. This variability is governed by the CDP, which allows a degree of disorder within set boundaries (31–41). Deviating from this range may result in malfunction, disease, and possibly malignancies (42). Chronic medication administration triggers compensatory responses, diminishing effectiveness in up to half of patients with chronic illnesses, which may be associated with steady dosing (13). Research is advancing in using variability-based algorithms to counteract the compensatory responses to chronic medications (12, 21–23, 40, 42–56). These CDP-based AI systems personalize treatments by dynamically adjusting dosages and timing according to patient-specific and disease-specific variables, aiming to improve drug effectiveness (13–15, 18). Machine learning is also being explored to refine treatments by leveraging personalized variability, further enhancing chronic medication efficacy (12, 14, 21, 22, 42–62).

In the first level of this methodology, as utilized in the current study, lenvatinib dosages and administration times were randomized within the clinically approved ranges through a pseudo-random number generator, creating an open-loop system that functions independently of direct feedback (13–16). In the upcoming levels, the algorithm adapts treatments based on patient responses and clinical outcomes, evolving into a closed-loop system that personalizes therapy by learning from aggregated user data (15). Additionally, incorporating chronotherapy and biological variability markers represents a further refinement, aiming to align treatment regimens more closely with individual pathophysiology (15).

Lenvatinib gained approval for locally recurrent or metastatic progressive RR-DTC following the pivotal SELECT trial - a phase III, double-blind study that showed improved progression-free survival with lenvatinib versus placebo (63). Lenvatinib-related AEs occur in nearly all patients, with the majority being grade 3 or higher, especially in older patients (6, 64). In the SELECT trial, 88.7% of patients older than 65 years exhibited grade 3 or higher AEs, most commonly hypertension and proteinuria (64). Several retrospective analyses assessing the real-life treatment patterns, adherence, and AEs of lenvatinib in RR-DTC also showed a significant grade ≥3 AEs incidence (65–67). The absence of self-reported AEs by patients 1, 3, and 4 may be associated with alterations in the AEs perception after long-term pre-enrollment lenvatinib administration (68). Most guidelines advocate for an initial high dose of 24 mg/day of lenvatinib for thyroid cancer. However, the elderly and frail patients may struggle with compliance or may not be suitable candidates for high doses due to anticipated toxicities (5, 6, 64). In the SELECT trial, despite an initially planned dose of 24 mg/day, 67.8% of patients receiving lenvatinib required dose reductions due to AEs, resulting in a mean dose of 17.2 mg/day. Additionally, 82.4% experienced dose interruptions, and 14.2% discontinued treatment due to toxicity, with even higher rates noted in older patients (63, 64). Other studies have reported substantial dose reductions or interruptions exceeding 80% within a few weeks of treatment initiation (6). These interruptions may affect disease progression and outcomes, as more prolonged disruptions are associated with a shorter progression-free survival (69).

Retrospective studies have proposed planned drug holidays as a strategy to address intolerable lenvatinib AEs and prevent recurrent treatment interruptions and potential tumor regrowth (70, 71). However, planned drug holidays are not routinely implemented in clinical practice for RR-DTC treatment, primarily due to concerns regarding disease progression (72). Treatment interruptions are undesirable and can be caused by side effects. Interruptions may lead to loss of effect of the drugs.

Therefore, a method for reducing drug interruptions while maintaining sufficient dosing is needed. In our trial, patients experienced dose interruptions of varying frequencies and durations. Pauses in treatment that are randomly introduced are not the same duration as a typical “planned drug holiday.” One patient discontinued treatment due to disease progression, with no discontinuations attributed to AEs. We attribute reliable drug administration reporting via the application’s user-friendly online system to reflect real-life adherence to the treatment regimen.

Despite the limitations of being an open-label design and small sample size, this study introduces a novel treatment strategy for patients receiving lenvatinib. The study’s relatively short duration precludes assessment of the algorithm’s long-term impact. Additionally, using the app as a reminder can enhance patients’ adherence and outcomes, irrespective of the variability mechanism (20).

In conclusion, conventional fixed regimens fail to address the complexity of lenvatinib treatment, marked by a high incidence of adverse events, emerging resistance mechanisms, and limited alternative therapies. The CDP-based AI-assisted approach offers a user-friendly framework for developing and implementing innovative treatment strategies, benefiting non-responsive patients who cannot tolerate maximum doses. We advocate alternating dose reductions and administration times rather than interruptions or drug holidays. Further research is warranted to assess the impact of variability-based treatment regimens, integrated with AI and digital tools, in reducing dosing interruptions, managing adverse events, and enhancing tumor response. The system provides a novel option for overcoming drug resistance in patients with malignancy.

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

The studies involving humans were approved by Hebrew University Hadassah IRB. The studies were conducted in accordance with the local legislation and institutional requirements. The participants provided their written informed consent to participate in this study.

YI: Conceptualization, Writing – original draft, Writing – review & editing. TS: Data curation, Investigation, Writing – original draft, Writing – review & editing. RG: Investigation, Writing – review & editing. OM: Conceptualization, Investigation, Writing – review & editing. AY: Data curation, Investigation, Writing – review & editing. NH: Data curation, Investigation, Writing – review & editing. SA: Data curation, Investigation, Writing – review & editing. MB: Conceptualization, Software, Writing – review & editing. AP: Conceptualization, Investigation, Writing – review & editing.