Following on from our blog on why causal thinking really matters in healthcare, we’re going to dive into some practical examples. Causal Machine Learning (Causal ML) is starting to gain adoption in many fields, and its application in healthcare, in particular surgery, is exciting and relatively new. Here are three pioneering studies that show how Causal ML can improve surgical care.
Uncovering True Causes of Length-of-Stay
Lee et al. (2022) – Cardiac Surgery Study
The Problem
Traditional statistical models can identify factors associated with longer hospital stays, but they don’t tell us what actually causes prolonged stays after cardiac surgery.
The Innovation
The researchers used a technique called Fast Causal Inference (FCI) to map out cause-and-effect relationships between different factors, creating causal graphs. It’s like drawing a map that doesn’t just show which events happen together, but which ones have causal influence on others.
Key Findings
- Clinicians’ intuitions about causes were largely correct
- Traditional statistical associations often differed from true causal relationships in both strength and direction
- Some factors previously thought important were actually just correlations, not causes
Practical Implications
This study helps clinical teams focus interventions on factors that genuinely affect length-of-stay, rather than wasting resources on factors that merely correlate with it.
Preventing Hospital Readmissions
Marafino et al. (2020) – Kaiser Permanente Study
The Problem
Traditional models predict who’s likely to be readmitted, but don’t tell us whose readmission we can actually prevent.
The Innovation
Using a method called causal forests1, the researchers developed a framework to identify not just high-risk patients, but those most likely to benefit from readmission prevention programs i.e. preventable risk. Then, they estimated the effect of applying the re-admission prevention program (the treatment) or not. It’s a type of ‘patient simulator’ to explore different ‘what-if’ scenarios for specific real patients.
Key Finding
- Targeting interventions based on estimated causal effect, prevented more readmissions than traditional risk-based approaches; found by comparing results before and after implementation
Practical Implications
- Hospitals can allocate limited resources more effectively by targeting patients whose outcomes are most likely to be improved
- Care teams can simulate the impact of interventions before applying them, enabling more informed decision-making
Limitations
The study focused on a specific readmission prevention program, making it less generalisable to hospitals without similar programs. The causal forest does not involve explicit modelling of the causal graph (the structure of causal relationships between variables). Instead, it makes assumptions about the causal structure, which could introduce bias.
Predicting Surgery Duration
Babayoff et al. (2022) – Sourasky Hospital Study
The Problem
Accurate prediction of surgery duration could dramatically improve scheduling, but traditional predictive models may focus on the wrong factors.
The Innovation
The researchers combined predictive modeling with causal analysis to improve the predictive models and identify truly influential factors for surgery duration. They treated all variables as ‘treatments’, so the work is applicable to any centre (even if they do not have a treatment such as the readmission prevention program in Marafino et. al 2020). They used the model to perform ‘what-if’ scenario planning at the population level, which could be helpful for ‘policy’ level decisions.
Key Findings
- Most features that were important for predictions did not have a strong causal relationship with surgery duration (also found by Lee et al. 2022 above)
- A streamlined model using only causal factors performed nearly as well as more complex models (but is likely to generalise better, although the researchers did not test that)
Practical Implications
Hospitals can improve scheduling by focusing on factors that actually influence surgery duration, rather than mere correlations.
Limitations
Like in Marafino et al., the researchers did not explicitly model the causal graph, which can hide assumptions that introduce bias.
Lessons Learned and Future Directions
Causal ML is relatively new for surgery. These early studies demonstrate the potential benefits and highlight three key insights:
- Causal effect matters: the most predictive features are not necessarily the ones with the potential to cause change; knowing what truly causes outcomes helps target interventions
- ‘What-if’ scenarios are powerful: estimating the effect of treatment alternatives helps make effective decisions
- Less can be more: models using fewer, but causally relevant features, can be as effective as complex ones
Future Challenges
- Explicitly modelling causal graphs of surgical settings to ensure that the causal ML model used is appropriate (and does not make incorrect assumptions that introduce bias)
- Validating causal models across different healthcare settings
- Translating this research into tools for real-time decision support
It’s really exciting to see researchers start to apply causal ML to surgical care. With just three pioneering examples, we already see diverse approaches and benefits. It’s our collective responsibility to consider causal thinking in all AI in healthcare to achieve the best results for patients. We are doing just that in the area of surgical decision-making. Stay tuned to see future updates.
Bibliography
Marafino BJ, Schuler A, Liu V, Escobar G, Baiocchi M. A Causal Machine Learning Framework for Predicting Preventable Hospital Readmissions. arXiv: Applications.
Babayoff O, Shehory O, Shahoha M, Sasportas R, Weiss-Meilik A. Surgery duration: Optimized prediction and causality analysis. PLOS ONE. 2022;17(8):e0273831. doi:10.1371/JOURNAL.PONE.0273831
Lee JJR, Srinivasan R, Ong CS, et al. Causal determinants of postoperative length of stay in cardiac surgery using causal graphical learning. The Journal of Thoracic and Cardiovascular Surgery. Published online August 27, 2022. doi:10.1016/j.jtcvs.2022.08.012
- Causal forests are an extension of random forests, a popular ML algorithm. ↩︎