Whose Labor Should Be Induced, and When? The Bishop Score Is a Guess. AI Can Do Better.
An article in "The Future of ObGyn" series
The Bishop score was published in 1964. It has five subjective variables, less than 50% specificity, and no awareness of the patient attached to the cervix. Machine learning models trained on real induction data already outperform it. We built a prototype to show how.
Post in The Future of ObGyn series.
A 30-year-old nulliparous woman presents at 39 weeks for elective induction. Her Bishop score is 4. The nurse starts a Foley balloon. Down the hall, another 30-year-old nulliparous woman has the same Bishop score. She gets misoprostol. Same hospital. Same indication. Same cervix. Different method, different tachysystole risk, different probability of vaginal delivery. Neither patient was told why one method was chosen over the other.
The decision about who should be induced, when, and how is arguably the most consequential routine decision in obstetrics. It affects more than one million American women per year. And we are making it with a scoring system from 1964 that treats the cervix as the entire patient.
The Bishop Score: 60 Years and Counting
Edward Bishop published his pelvic scoring system in Obstetrics & Gynecology in 1964. It assigns points for five cervical characteristics: dilation, effacement, station, consistency, and position. A score of 6 or higher is considered “favorable” for induction. Below 6 is “unfavorable.”
The score has one virtue: simplicity. It also has fundamental problems.
First, it is subjective. Two examiners assessing the same cervix will disagree on effacement and consistency a significant proportion of the time. Second, its predictive performance is poor. A 2025 systematic review in Cureus confirmed that the Bishop score has sensitivity around 60% and specificity below 50% for predicting vaginal delivery after induction. Third, and most important, the Bishop score has no variables for the patient herself: her parity, her BMI, her height, her prior deliveries, her fetal weight, or her reason for induction. It treats the cervix as if it exists in isolation.
The strongest single predictor of successful induction is not cervical dilation. It is parity. A parous woman with a Bishop score of 3 has a higher probability of vaginal delivery than a nulliparous woman with a Bishop score of 7. The Bishop score cannot express this.
Machine Learning Already Does Better
Several groups have now published machine learning models that significantly outperform the Bishop score for predicting induction outcomes. Ferreira and colleagues (Acta Obstet Gynecol Scand, 2025) developed a multivariable ML model using SHAP (SHapley Additive exPlanations) to identify which variables actually drive the prediction. Their model achieved excellent discrimination. Zhang and colleagues (Scientific Reports, 2022) tested four different ML algorithms on 907 induction cases and found all four outperformed traditional scoring. Krsman and colleagues (European Review, 2023) added ultrasound cervical assessment to clinical variables and achieved 83% accuracy.
These are not theoretical exercises. They are working models, trained on real patient data, that can tell a clinician: this specific woman has a 78% probability of vaginal delivery with this induction method, a 15% tachysystole risk, and would be better served by a Foley balloon than by misoprostol.
The question is not whether AI can do better than the Bishop score. The question is why we are still using the Bishop score at all.
🎯 Free Subscriber Bottom Line: The Bishop score, published in 1964, remains the standard tool for predicting induction success despite sensitivity of 60% and specificity below 50%. It ignores parity, BMI, fetal weight, prior delivery mode, and indication for induction. Machine learning models incorporating these variables already achieve over 80% accuracy. An AI-based induction decision support tool could tell clinicians and patients the probability of vaginal delivery, the tachysystole risk, and the optimal method for each individual patient. We built a working prototype to demonstrate what this looks like.
Below, paid subscribers get:
A suggested INTERACTIVE AI model to help decide if and when to induce. A tool you can use TODAY.
The ML models that outperform the Bishop score: Ferreira 2025, Zhang 2022, Krsman 2023, and what they found
2. Why the ARRIVE trial does NOT support elective induction in multiparous women
3. Evidence-based delivery timing by indication: what ACOG says, what the evidence actually supports
4. How induction method changes your probability: balloon vs misoprostol vs oxytocin vs amniotomy
5. The AI Induction Decision Support Tool: a working interactive prototype you can use right now [LINK]
6. The full evidence base: 18 publications informing the model
7. What a validated tool would need: the path from prototype to bedside
PAID SUBSCRIBER CONTENT
