Development and internal validation of a pre-operative machine-learning model for prolonged postoperative stay in older adults undergoing colorectal cancer surgery
This study developed and internally validated a pre-operative machine-learning model using routine clinical data to moderately predict prolonged hospital stays in older adults undergoing colorectal cancer surgery, identifying frailty and comorbidity as key predictors while highlighting the need for external validation before clinical implementation.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Hospitals are places of healing, but for older adults, a long stay can become a source of new harm. When an elderly person is admitted for major surgery, the hospital environment itself can trigger a decline in their physical and mental abilities. This phenomenon, known as hospital-associated disability, includes a loss of muscle strength, confusion, and a reduced ability to care for oneself, all of which can happen simply because a patient is confined to a bed or a room for too long. For patients facing colorectal cancer surgery, the challenge is compounded by the fact that they often arrive with a mix of other health issues, reduced physical reserve, and complex social needs that make returning home difficult. Doctors have long known that frailty—a state of increased vulnerability to stress—is a better predictor of poor outcomes than age alone, yet predicting exactly who will need a long hospital stay remains difficult. Traditional tools often focus on the risk of death or major complications, missing the subtler but equally damaging risk of a prolonged admission that leaves a patient weaker than when they arrived.
In a recent study, researchers at Southend University Hospital and Southmead Hospital set out to build a digital tool to solve this specific problem. They wanted to know if they could use a computer program to look at a patient's health data before surgery and predict whether that patient would stay in the hospital for more than ten days. The team focused on 197 older adults, all aged 65 or older, who were being assessed for colorectal cancer surgery. From this group, they analyzed the records of 149 patients who actually underwent the operation. The researchers gathered a wide range of information available before the surgery began, including the patient's age, a measure of their frailty called the Clinical Frailty Scale, the number of other medical conditions they had, how many medications they took, and their social situation, such as where they lived and what help they had at home. They also included standard surgical risk scores that estimate the difficulty of the operation.
To make sense of this complex mix of data, the team used a machine-learning method called a gradient-boosted classifier. You can think of this as a computer program that learns by building a series of decision trees, where each new tree corrects the mistakes of the one before it, eventually creating a highly accurate model for spotting patterns. The researchers trained this model to distinguish between patients who would have a short or moderate stay and those who would have a prolonged stay of more than ten days. They tested the model rigorously by repeatedly splitting the data into different groups to ensure the results were not just a lucky guess. The model proved moderately successful at its task. It correctly identified patients at risk of a long stay about 75% of the time, with a sensitivity of 66%, meaning it caught two-thirds of the people who actually ended up staying too long. It was also reasonably good at ruling out those who would have a shorter stay, with a specificity of 75%.
The study revealed that the most powerful clues for predicting a long stay were not simply the patient's age or the type of surgery, but rather their level of frailty and their overall surgical risk score. The number of other health conditions and the patient's mobility also played significant roles. Interestingly, the researchers found that while the model was decent at predicting a binary outcome—whether a stay would be long or not—it was much less reliable at predicting the exact number of days a patient would spend in the hospital. When they tried to use the computer to forecast specific durations, the predictions were often off by several days, suggesting that the exact length of a hospital stay is influenced by too many unpredictable factors to be calculated precisely in advance. This distinction is crucial: the tool is better at flagging a high-risk patient who needs extra planning than it is at telling a doctor exactly when that patient will be discharged.
The implications of being able to identify these high-risk patients are significant. The researchers estimated that if this model were applied to the roughly 21,000 major bowel cancer surgeries performed annually across England and Wales, it could help identify thousands of patients who are likely to face a prolonged stay. A prolonged admission is not just a matter of hospital bed availability; it is a direct threat to the patient's independence. The study highlighted that patients who stay longer are exposed to higher rates of functional decline, delirium, and other hospital-acquired complications. By identifying these patients early, medical teams could potentially intervene with more intensive pre-operative preparation or post-operative support to shorten their stay. The economic argument is also compelling; the researchers calculated that reducing the length of stay by just one day for 10% of these national cases could save nearly £2 million in bed costs alone, a figure that would rise to over £5 million if the reduction were achieved in 20% of cases.
Despite these promising results, the authors are careful to note that this tool is not yet ready for immediate use in every hospital. The study was conducted at a single center and relied on retrospective data, meaning the model was tested on past records rather than future patients. The performance of the model varied depending on how the data was tested, and the researchers emphasized that it needs to be validated in a larger, more diverse group of patients before it can be trusted to guide clinical decisions. The study concludes that while machine learning shows great promise for helping doctors plan care for older adults undergoing colorectal cancer surgery, the next step is to test the model in real-time, prospective trials to see if acting on its predictions actually leads to shorter, safer hospital stays. Until then, the tool remains a sophisticated aid for risk stratification, offering a clearer view of who is most vulnerable to the hidden dangers of a long hospital admission.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.