In the dawn of artificial intelligence’s potential to inform clinical practice, the importance of understanding the intent and interpretation of prediction tools is vital. In medicine, informed decision-making promotes patient autonomy and can lead to improved patient satisfaction and engagement in their own care.
This shared understanding enables patients to make informed choices about their care, reducing anxiety and increasing confidence in medical decision-making.
Tufts University
Dr. Sebastian Z. Ramos
In obstetric clinical practice, prediction tools have been created to assess risk of primary cesarean delivery in gestational diabetes,1 cesarean delivery in hypertensive disorders of pregnancy,2 and failed induction of labor in nulliparous patients with an unfavorable cervix.3 By assessing a patient’s risk profile, clinicians can identify high-risk individuals who may require closer monitoring, early interventions, or specialized care. This allows for more timely interventions to optimize maternal and fetal health outcomes.
Other prediction tools are created to better elucidate to patients their individual risk of an outcome that may be modifiable, aiding physician counseling on mitigating factors to improve overall results. A relevant example is the American Diabetes Association’s risk of type 2 diabetes calculator used for counseling patients on risk reduction. This model includes both preexisting (ethnicity, family history, age, sex assigned at birth) and modifiable risk factors (body mass index, hypertension, physical activity) to predict risk of type 2 diabetes and is widely used in clinical practice to encourage integration of lifestyle changes to decrease risk.4 This model highlights the utility of prediction tools in counseling, providing quantitative data to clinicians to discuss a patient’s individual risk and how to mitigate that risk.
While predictive models clearly have many advantages and potential to improve personalized medicine, concerns have been raised that their interpretation and application can sometimes have unintended consequences as the complexity of these models can lead to variation in understanding among clinicians that impact decision-making. Different clinicians may assign different levels of importance to the predicted risks, resulting in differences in treatment plans and interventions. This variability can lead to disparities in care and outcomes, as patients with similar risk profiles may receive different management approaches based on the interpreting clinician.
Providers may either overly rely on prediction models or completely disregard them, depending on their level of trust or skepticism. Overreliance on prediction models may lead to the neglect of important clinical information or intuition, while disregarding the models may result in missed opportunities for early intervention or appropriate risk stratification. Achieving a balance between clinical judgment and the use of prediction models is crucial for optimal decision-making.
An example of how misinterpretation of the role of prediction tools in patient counseling can have far reaching consequences is the vaginal birth after cesarean (VBAC) calculator where race and ethnicity naturalized racial differences and likely contributed to cesarean overuse in Black pregnant people as non-White race was associated with a decreased chance of successful VBAC. Although the authors of the study that created the VBAC calculator intended it to be used as an adjunct to counseling, institutions and providers used low calculator scores to discourage or prohibit pregnant people from attempting a trial of labor after cesarean (TOLAC). This highlighted the importance of contextualizing the intent of prediction models within the broader clinical setting and individual patient circumstances and preferences.
This gap between intent and interpretation and subsequent application is influenced by individual clinician experience, training, personal biases, and subjective judgment. These subjective elements can introduce inconsistencies and variability in the utilization of prediction tools, leading to potential discrepancies in patient care. Inadequate understanding of prediction models and their statistical concepts can contribute to misinterpretation. It is this bias that prevents prediction models from serving their true purpose: to inform clinical decision-making, improve patient outcomes, and optimize resource allocation.
Clinicians may struggle with concepts such as predictive accuracy, overfitting, calibration, and external validation. Educational initiatives and enhanced training in statistical literacy can empower clinicians to better comprehend and apply prediction models in their practice. Researchers should make it clear that models should not be used in isolation, but rather integrated with clinical expertise and patient preferences. Understanding the limitations of prediction models and incorporating additional clinical information is essential.
Prediction models in obstetrics should undergo continuous evaluation and improvement to enhance their reliability and applicability. Regular updates, external validation, and recalibration are necessary to account for evolving clinical practices, changes in patient populations, and emerging evidence. Engaging clinicians in the evaluation process can foster ownership and promote a sense of trust in the models.
As machine learning and artificial intelligence improve the accuracy of prediction models, there is potential to revolutionize obstetric care by enabling more accurate individualized risk assessment and decision-making. Machine learning has the potential to significantly enhance prediction models in obstetrics by leveraging complex algorithms and advanced computational techniques. However, the unpredictable nature of clinician interpretation poses challenges to the effective utilization of these models.
By emphasizing communication, collaboration, education, and continuous evaluation, we can bridge the gap between prediction models and clinician interpretation that optimizes their use. This concerted effort will ultimately lead to improved patient care, enhanced clinical outcomes, and a more harmonious integration of these tools into obstetric practice.
Dr. Ramos is assistant professor of maternal fetal medicine and associate principal investigator at the Mother Infant Research Institute, Tufts University and Tufts Medical Center, Boston.
References
1. Ramos SZ et al. Predicting primary cesarean delivery in pregnancies complicated by gestational diabetes mellitus. Am J Obstet Gynecol. 2023 Jun 7;S0002-9378(23)00371-X. doi: 10.1016/j.ajog.2023.06.002.
2. Beninati MJ et al. Prediction model for vaginal birth after induction of labor in women with hypertensive disorders of pregnancy. Obstet Gynecol. 2020 Aug;136(2):402-410. doi: 10.1097/AOG.0000000000003938.
3. Levine LD et al. A validated calculator to estimate risk of cesarean after an induction of labor with an unfavorable cervix. Am J Obstet Gynecol. 2018 Feb;218(2):254.e1-254.e7. doi: 10.1016/j.ajog.2017.11.603.
4. American Diabetes Association. Our 60-Second Type 2 Diabetes Risk Test.