Clinical Decision Support Models to Optimize Controlled Ovarian Hyperstimulation: A Scoping Review

All authors wrote the manuscript and designed the research; R.Z.Y.L., T.J.F.M., and S.L.S. performed the research; R.Z.Y.L., T.J.F.M., and S.L.S. analyzed the data.

The authors have nothing to report.

This scoping review was conducted in accordance with the Preferred Reporting Items for Systematic reviews and Meta‐analyses extension for scoping reviews (PRISMA‐ScR) [ 9 ]. A search was conducted in MEDLINE, Embase, PubMed, and Scopus from database inception to 24 March 2026. Search strategies combined keywords related to ovarian stimulation, clinical decision support systems, and predictive modeling (including machine learning and artificial intelligence). Search strategies were adapted for each database and developed with an experienced academic librarian (Tess Aitken, The University of Sydney) to maximize sensitivity and specificity. The full search strategies are provided in Figure  S1 . Reference lists of included studies were also screened to identify any additional relevant publications. Studies were eligible for inclusion if they were published in English and investigated a CDS tool or predictive model intended to inform precision dosing in COH, applied model‐based, ML, or AI approaches, evaluated the use of the model in a clinical/clinically simulated setting to inform gonadotropin dosing, prediction of ovarian response and/or timing of ovulation trigger in women undergoing ART procedures. Studies were required to report on the number of oocytes/metaphase II (MII) oocytes retrieved and include information on model training and validation. Studies that developed prediction models without evaluating their application in decision‐making were excluded. Secondary literature including secondary analysis of included studies, reviews, case reports, editorials, letters, and book chapters were excluded. The primary outcome was the number of oocytes/MII oocytes retrieved per initiated COH cycle. Secondary outcomes included clinical pregnancy rate, ongoing pregnancy rate, live birth rate, starting FSH dose, and total FSH dose administered. Additional data extracted included study and population characteristics, model type (e.g., AI, ML, PK/PD), model development and validation methods, input variables (e.g., age, body mass index (BMI), AMH, AFC), and model performance metrics (e.g., mean average error (MAE), coefficient of determination ( R 2 )). The intended clinical use of the model (e.g., dosing recommendation, trigger timing decision) was also recorded. Search results were imported into Covidence [ 10 ] and duplicates were removed. Title and abstract screening were conducted independently by two reviewers (RL, TM). Any conflicts were deferred to a third independent reviewer (SS). The full texts of potentially eligible articles were then reviewed, with discrepancies resolved by a third reviewer (SS). Relevant data were extracted using a standardized data collection form. Extraction was performed (RL) and independently verified (TM) for accuracy and completeness. For studies reporting data graphically, numerical values were estimated using PlotDigitizer [ 11 ]. Where outcomes were reported using medians and ranges, the mean and standard deviation were estimated using established conversion methods [ 12 ]. Such conversions were annotated with (C). In line with the PRISMA‐ScR framework [ 9 ], formal risk of bias assessment of individual studies was not undertaken. However, methodological quality and completeness of reporting were considered by both reviewers during data extraction and synthesis to inform interpretation of findings. A descriptive synthesis was undertaken due to heterogeneity in study designs, model types, and reported outcomes. Studies were grouped according to the primary predictive domain of the model: gonadotropin dosing (precision dosing), timing of ovulation trigger, or combined dosing and trigger timing approaches. Where available, individual studies' own reported effect estimates and significance values were presented. Comparisons focused on differences in oocyte yield and gonadotropin use between model‐informed and standard care approaches. Model performance metrics and input variables were summarized to identify common methodological approaches and predictors.

From 2879 identified records, five studies were included (Figure  1 ), comprising 16,441 participants (Table  S1 ). The studies were published between 2022 and 2025. The weighted mean age was 36.05 years, although reporting of variability measures was inconsistent. Study characteristics are presented in Table  1 . PRISMA Diagram. Characteristics and Demographics. PCOS: 6.95 Endometriosis: 4.05 Tubal factors: 15.05 Treatment: 292 Control: 292 Treatment: 291 Control: 291 Treatment: 36.10 Control: 36.07 Treatment: 24.66 Control: 25.68 Treatment: 2.63 Control: 2.50 Treatment: 17.20 Control: 16.47 Note: Data reported as mean (SD) or n (%), where reported. Abbreviations: AFC: antral follicle count; AMH: anti‐Müllerian hormone; bFSH: baseline follicle stimulating hormone; C: calculated; NR: not reported. Evaluated the same clinical decision support software (Stim Assist), but applied to distinct patient cohorts at different clinical sites. Three studies used AI‐based prediction models [ 13 , 14 , 15 ] and two used ML‐based models [ 16 , 17 ]. No clinically evaluated mechanistic PK/PD or quantitative systems pharmacology approaches were identified. The CDS tools were categorized according to their clinical function. Two models focused on precision‐dosing, which aimed to individualize gonadotropin starting dose [ 13 , 16 ]. One focused on ovulation trigger timing, which aimed to optimize the timing of oocyte retrieval [ 17 ]. Two combined dosing and trigger decision support [ 14 , 15 ]. Among the four precision‐dosing algorithms, commonly included input variables were age, BMI, AFC, and AMH (Table  2 ). The normalized mean impact value of these key predictors was 1, 0.439, −0.499, −0.841 respectively [ 13 ]. Models with expanded feature sets additionally incorporated endocrine parameters (e.g., baseline FSH, luteinizing hormone, oestradiol), infertility characteristics, and prior treatment history [ 13 ]. Intervention. Note: Data reported as mean (SD) or n (%), where reported. Abbreviations: AFC, antral follicle count; AMH, anti‐Müllerian hormone; ANN, artificial neural network; BMI, body mass index; C, calculated; SD, standard deviation; E2, oestradiol; FSH, follicle stimulating hormone; GSD, gonadotropin starting dose; IVF, in vitro fertilization; KNN, K‐nearest neighbors; LH, luteinizing hormone; MAE, mean absolute error; MII, metaphase II; NR, not reported; R 2 , coefficient of determination; RMSE, root mean square error. Evaluated the same clinical decision support software (Stim Assist), but applied to distinct patient cohorts at different clinical sites. Ovulation‐trigger timing models ( n  = 3) relied on follicular size and serum oestradiol concentrations, reflecting their direct relationship with follicular maturation [ 14 , 15 , 17 ]. Among studies evaluating precision dosing ( n  = 4), K‐nearest neighbors (KNN) algorithms predominated ( n  = 3) [ 14 , 15 , 16 ], while one study implemented an artificial neural network (ANN) approach [ 13 ]. All studies examining ovulation‐trigger timing ( n  = 3) used linear regression [ 14 , 15 , 17 ]. Two studies evaluated CDS‐guided gonadotropin dosing strategies against clinician‐directed dosing [ 13 , 16 ]. One study reported that model‐informed dosing was associated with an increase in the number of MII oocytes retrieved (+1.5 oocytes; p  < 0.001), and a decrease in the total gonadotropin dose (−195 IU; p  < 0.001) [ 16 ]. One study assessed a CDS tool designed to optimize ovulation trigger timing. Compared with standard care, model‐guided trigger timing increased MII oocyte yield by 2.5 oocytes ( p  < 0.01). Two studies evaluated integrated CDS tools that informed both gonadotropin dosing and ovulation trigger timing. One study reported a non‐significant increase in MII oocytes (+0.96 oocytes; p  = 0.16) and a non‐significant reduction in gonadotropin use (−174.35 IU; p  = 0.13). The other study reported no difference in MII oocyte yield (−0.07 oocytes; p  = 0.90), but a significant decrease in total gonadotropin dose (−472.94 IU; p  < 0.01). Across all studies, neither clinical pregnancy rates nor live birth outcomes were reported (Table  3 ). The number of MII oocytes was therefore used as a surrogate outcome for treatment success. Clinical Outcomes. NR treatment arm: 10.81 +/− 7.67 (C) Note: Data reported as mean (SD) or n (%), where reported. Abbreviations: C, calculated; FSH, follicle stimulating hormone; MII, metaphase II; NR, not reported; SD, standard deviation. Evaluated the same clinical decision support software (Stim Assist), but applied to distinct patient cohorts at different clinical sites. Quantitative model performance metrics were reported inconsistently. For ovulation trigger timing and gonadotropin dosing, performance metrics from two studies are detailed in Table  S2 . Notably, prediction of next‐day oestradiol concentrations achieved higher accuracy than MII oocyte prediction ( R 2  = 0.88 vs. 0.62–0.64). Other studies focused primarily on clinical outcomes and provided limited information regarding discrimination, calibration, or external validation. One precision dosing study [ 13 ] directly compared model‐recommended gonadotropin starting doses with clinician‐selected doses. Clinicians were more likely to prescribe doses at the extremes of the dosing range, whereas model predictions were more centralized. The magnitude of discrepancy between clinician‐ and model‐recommended doses increased at the low and high ends of the oocyte retrieval distribution, suggesting greater divergence in predicted versus empirical decision‐making in patients expected to have either low or high ovarian response. Safety reporting was limited. Two studies [ 14 , 15 ] reported no cases of OHSS, whilst the remaining studies did not report OHSS outcomes. As such, the available evidence is insufficient to determine the safety implications of CDS‐guided COH.

This scoping review synthesized the current evidence on CDS tools designed to inform decision‐making during COH. Across five heterogeneous studies involving 16,441 patients, model‐informed approaches were applied to gonadotropin dosing, ovulation trigger timing, or both. Overall, model‐informed dosing may improve mature oocyte yield while reducing gonadotropin exposure, suggesting that precision medicine strategies may improve the efficiency of ovarian hyperstimulation. However, the evidence base remains limited by small numbers of studies, reliance on surrogate outcomes, and a lack of mechanistic modeling approaches. CDS‐guided gonadotropin dosing consistently improved oocyte yield while reducing gonadotropin exposure. The biological plausibility of precision dosing in COH is supported by variability in follicular sensitivity to FSH [ 5 ]. Uniform dosing strategies may lead to heterogenous responses to exogenous stimulation within a cohort due to inherent differences in FSH receptor expression and downstream signaling pathways [ 18 ]. Individualized dosing aims to align FSH exposure with patient‐specific ovarian sensitivity, thereby improving recruitment of responsive follicles whilst limiting excessive stimulation [ 5 ]. The mechanism may also explain the consistent reduction in total gonadotropin‐use observed in CDS‐guided dosing strategies. Patients with higher intrinsic sensitivity may benefit from dose reductions, whilst escalations of predicted poor responders appeared to have limited impact on improving oocyte yield. The observed reduction in gonadotropin exposure is particularly notable. Effective precision dosing should improve therapeutic efficiency by achieving comparable or improved outcomes with lower drug utilization. CDS‐guided dosing may move clinical practice closer to this objective, supporting the broader concept of MIPD within reproductive medicine. Although one trigger‐timing model demonstrated an improvement in MII oocyte yield, studies evaluating combined dosing‐and‐trigger approaches did not consistently outperform dosing strategies alone. Optimisation of follicular development through appropriate gonadotropin dosing may represent the primary determinant of treatment success, while the role of trigger‐timing algorithms may be more context‐dependent. One possible explanation is that individualized dosing promotes greater follicular synchrony. Follicles exhibit intrinsic variability in FSH receptor density and isoform composition, producing a spectrum of sensitivities to stimulation [ 18 ]. As such, each follicle requires a different amount of FSH to achieve the cAMP/PKA activation threshold that initiates granulosa‐cell proliferation [ 18 ]. Therefore, individualized FSH dosing may promote more uniform follicular recruitment and growth, thereby reducing inter‐follicular variability at the time of trigger [ 18 ]. Therefore, the need for precise trigger timing may be diminished as a more homogenous follicular cohort is less sensitive to small differences in timing. While this is biologically plausible, the current evidence base remains insufficient to evaluate this mechanism. Homogenous growth of follicles and reduced total dose of FSH used is consistent with previous findings. A case–control study found individualized dosing markedly lowered the incidence of early and late follicular asynchrony relative to a clinician‐based control group (4.5% vs. 17.0% on Day 5 of COH; 2.3% vs. 37.5% on Day 9 of COH) [ 19 ]. Similarly, the study also found a lower total FSH dose used (−822 IU, p  < 0.001) [ 19 ]. Another retrospective analysis shows that providing an adequate FSH dose from the outset of stimulation reduces inter‐follicular size dispersion and enhances mature oocyte yield [ 20 ]. This has meaningful implications for the next generation of CDS tools, indicating precision dosing algorithms alone may benefit women undergoing IVF and do not require additional trigger‐timing modules, as this may contribute additional operational burden without improving outcomes. Across precision‐dosing models, there was strong convergence on model predictors: age, BMI, AFC, and AMH. Age captures the progressive depletion of FSH‐sensitive antral follicles, requiring higher exogenous FSH doses in older women [ 21 ]. AMH and AFC directly quantify the size and functional activity of the recruitable antral follicle pool, reflecting how much FSH is required to achieve multi‐follicular growth [ 21 ]. Excess body weight dilutes circulating FSH through increased extracellular fluid distribution [ 22 ], meaning BMI omission risks systematic dose misestimation, particularly in populations with high weight variability such as PCOS [ 22 ]. For trigger‐timing models, oestradiol concentration and follicular size were the dominant inputs. Oestradiol reflects granulosa cell activity and correlates positively with oocyte number and maturity [ 21 ], while follicular size on the day of trigger determines oocyte recovery [ 23 ]. Notably, more complex models incorporating additional variables did not provide substantial incremental predictive value beyond these core parameters, likely reflecting the strong physiological coupling between follicular growth and endocrine output in the late follicular phase [ 17 ]. The included studies used relatively simple and interpretable algorithms. KNN and linear regression methods accounted for most clinically evaluated tools, whereas more complex neural networks were uncommon, reflecting a broader challenge in implementation science. Although sophisticated black box models such as ANN algorithms capture complex nonlinear relationships, clinical adoption requires transparency, interpretability, and clinician trust [ 24 ]. KNN and linear regression models allow clinicians to understand the factors influencing a recommendation, potentially facilitating clinical integration [ 25 ]. The preference for interpretable approaches may therefore represent a pragmatic response to the requirements of real‐world implementations. Perhaps the most striking finding was the complete absence of clinically evaluated PK/PD or mechanistic modeling approaches among clinically evaluated CDS tools. In contrast to data‐driven ML and AI models, PK/PD frameworks provide a mechanistic representation of drug exposure, receptor interaction, and downstream biological response. Such mechanistic approaches have improved precision dosing in other therapeutic areas [ 8 ], yet they remain unexplored in ovarian hyperstimulation. Such approaches offer several advantages over purely data‐driven methods, including improved interpretability, extrapolation beyond observed datasets, and incorporation of known physiological mechanisms. The lack of mechanistic COH models represents an opportunity for future research. Follicular growth, endocrine responses, and gonadotropin exposure are all processes that may be amenable to mechanistic modeling. Integration of PK/PD principles with contemporary ML approaches could potentially produce hybrid CDS tools that combine physiological realism with predictive performance. The comparative evaluation of gonadotropin dosing tools, ovulation‐trigger timing tools, and combined models in one setting enables a structured understanding of how different CDS approaches address distinct phases of COH. By synthesizing the variables and algorithms used across studies, this review provides practical guidance for future developers regarding predictor selection and methodological design. This review should be interpreted in light of several limitations. First, the number of included studies was small ( n  = 5), and study designs, populations, and model structures were heterogeneous, limiting comparability; however, this heterogeneity also reflects the inherent variability of real‐world patient populations, suggesting our findings are applicable to clinical practice. As 4 of the 5 studies were conducted in the United States, the population‐level variations in ovarian responsiveness to COH [ 26 ] may also impact the generalisability of the findings of this review. Second, most studies were observational and lacked external validation or randomized comparisons, increasing the risk of bias. Third, outcome reporting was limited to surrogate measures, with no data on clinical pregnancy or live birth outcomes, meaning the clinical significance of the observed improvements remains uncertain. Additionally, model performance metrics were inconsistently reported, and safety outcomes were absent or inadequately described. The exclusion of model‐development‐only studies means that emerging approaches not yet evaluated in clinical settings were not captured. Despite encouraging findings, evidence supporting clinical implementation remains weak. No studies reported pregnancy or live birth outcomes, and safety reporting was sparse. Consequently, improvements in MII oocyte yield translating to meaning for patient benefits remain unclear. The need for future studies moving beyond surrogate endpoints and evaluating patient‐centred outcomes, implementation feasibility, clinician acceptance, economic evaluation, and accessibility. Such work will be essential if CDS tools are to progress from promising technologies to routinely deployed precision medicine interventions.

Current evidence suggests that CDS‐enabled precision gonadotropin dosing may improve oocyte yield while reducing drug exposure during COH. However, the field remains in an early stage of development and is currently dominated by empirical ML approaches evaluated against surrogate outcomes. The absence of mechanistic model‐informed approaches represents an important knowledge gap and a potential opportunity for innovation. Future research should include prospective evaluation of clinically meaningful outcomes and the development of hybrid ML‐mechanistic frameworks capable of supporting truly personalized COH management.

Infertility impacts one in six people globally [ 1 ], representing a growing healthcare challenge driven by biological and societal factors [ 2 ]. Assisted reproductive technologies (ART), such as in vitro fertilization (IVF), have enabled millions of individuals to have children. Despite this, treatment outcomes remain suboptimal. In 2018, only 24% of people who underwent oocyte retrieval globally achieved a live birth [ 3 ]. The financial, physical, and psychological burden [ 2 ] associated with repeated treatment cycles highlights the need for more effective treatment strategies. Controlled ovarian hyperstimulation (COH) is a critical component of ART and represents an important precision medicine challenge. During COH, exogenous gonadotropins, primarily follicle stimulating hormone (FSH), are administered to stimulate multi‐follicular development and maximize the number of mature oocytes available for fertilization [ 4 ]. Following sufficient follicular maturation, ovulation is triggered to release the oocytes [ 4 ]. The effectiveness of COH is dependent on both gonadotropin dosing and the timing of ovulation trigger administration, as these decisions influence oocyte yield, treatment efficiency, and safety [ 4 ]. A major challenge in COH is the substantial inter‐individual variability in ovarian response. Differences in ovarian reserve, age, endocrine function, body composition, and intrinsic follicular sensitivity can result in markedly different responses despite similar treatment regimens [ 5 ]. Current fixed or empiric dosing approaches may lead to under‐stimulation in some patients, reducing oocyte yield and treatment success, while causing excessive stimulation in others, increasing the risk of ovarian hyperstimulation syndrome (OHSS) [ 5 ]. Consequently, COH can be considered a precision dosing problem, where therapeutic decisions must balance efficacy and safety while accounting for considerable biological variability. Currently, gonadotropin dosing and ovulation trigger timing are guided by clinician experience combined with relatively simple predictors of ovarian reserve such as age, anti‐Müllerian hormone (AMH), and antral follicle count (AFC) [ 6 ]. While these factors provide valuable information regarding expected ovarian reserve, they do not fully capture the complex physiological processes underlying folliculogenesis. Consequently, substantial variability in treatment approaches and outcomes persists, highlighting the need for more quantitative and reproducible methods of clinical decision‐making. Clinical decision support (CDS) tools offer a framework to address this challenge. By integrating patient‐specific characteristics and treatment response data, CDS tools can generate personalized recommendations regarding gonadotropin dosing or ovulation trigger timing. These tools may utilize traditional statistical models, machine learning (ML), artificial intelligence (AI), or mechanistic pharmacokinetic/pharmacodynamic (PK/PD) approaches, conceptually aligning with model‐informed precision dosing (MIPD) in optimizing therapy through quantitative integration of patient data and predictive models. The application of ML and AI within reproductive medicine has expanded rapidly in recent years, particularly in embryo selection and estimation of pregnancy or live birth probabilities [ 7 ]. However, comparatively less attention has been directed toward the use of CDS tools to inform therapeutic decisions during COH, despite having direct influence on both treatment effectiveness and safety. Furthermore, while mechanistic model‐based approaches have demonstrated value in other areas of precision dosing, their role in reproductive medicine remains unclear [ 8 ]. Although reviews have examined the broader use of AI in ART, the evidence supporting CDS tools specifically designed to guide gonadotropin dosing, ovarian response management, and ovulation trigger timing has not been systematically synthesized. A comprehensive evaluation of these approaches is required to understand their current clinical utility, identify common modeling strategies and predictors, and determine key evidence gaps that may limit translation into clinical practice. Accordingly, this scoping review aimed to identify and evaluate CDS tools designed to inform prescribing decisions during COH. Particular focus was given to model characteristics, input variables, clinical outcomes, and how these tools support personalized and model‐informed treatment strategies in reproductive medicine.

The authors declare no conflicts of interest.

Table S1: Key terms used to build the search strategy using the PICO framework. Table S2: Model parameters. Figure S1: Search strategy used for each database.