Letter to the Editor: Machine learning models for hospital-acquired functional decline in geriatric surgical patients - from prediction to prevention

Abstract

Hospital-acquired functional decline (HAFD), even after a successful surgery, is found in up to one-third of the geriatric population. Early and accurate prediction of HAFD can provide timepoints for early intervention to improve postoperative outcomes. While machine learning (ML) offers a promising approach, several pitfalls must be addressed before clinical adoption. When combined with clinically interpretable outputs and continuous model recalibration, ML can support patient-centred decision-making, helping reduce preventable disability, shorten recovery time, and improve overall quality of care for vulnerable hospitalized elders.

Key Words: Frailty; Geriatric; Hospital-acquired functional decline; Machine learning; Prediction; Prevention

Core Tip: Machine learning (ML) for hospital-acquired functional decline will be truly successful not when it merely predicts decline, but when it enables timely interventions that prevent it. When combined with clinically interpretable outputs and continuous model recalibration, ML can support patient-centred decision-making, helping reduce preventable disability, shorten recovery time, and improve overall quality of care for vulnerable hospitalized elders.

TO THE EDITOR

The steadily increasing numbers of geriatric patients undergoing elective major surgery over past decades, has provided clinicians with new insights into their specialised concerns such as multiple comorbidities, polypharmacy, cognitive and mobility impairment, nutritional status, frailty and sarcopenia which may affect their postoperative outcomes[1-3]. Such patients increasingly prioritise meaningful postoperative functional recovery, dignity and independence over hospital discharge and survival alone. It is thus imperative that recovery in older adults be viewed as a continuum of care that begins at the time of diagnosis and extends well beyond hospital discharge, continuing until the patient regains functional capacity comparable to the premorbid state and successfully reintegrates into society[2]. While modern medicine has significantly improved survival from acute illness, a significant number of geriatric patients unfortunately leave the hospital with reduced mobility, impaired activities of daily living (ADL), a loss of independence, an increased need for caregiver support or institutional care and high healthcare costs[1]. Hospital-acquired functional decline (HAFD) can occur in patients of any age group following major surgery, prolonged bed rest, critical illness, or pain-related immobilization[4]. Geriatric patients with cardiovascular diseases, undergoing major cardiac surgery (with or without cardiopulmonary bypass) are at a significantly higher risk of HAFD[5]. Unlike outcomes such as mortality, intensive care unit admission, or readmission, functional decline is multifactorial and often silent in its early phase. Even, mild cognitive impairment combined with poor nutrition and bed rest can accelerate loss of function. Since the effect on postoperative functional decline increases exponentially with rising age from 60 years to 80 years and maybe much stronger in patients with frailty, HAFD represents a silent epidemic, where even after a successful surgery, up to one-third of the geriatric population experience significant functional decline during and after their hospital stay[1,4]. There is a growing body of literature that explores HAFD risk factors and definitions but definitive predictive tools to aid diagnosis still elude us[6,7]. Early and accurate prediction of HAFD has the potential to provide timepoints for early intervention to improve postoperative outcomes. For clinicians, early prediction and diagnosis offers a vantage point by changing their role; from simply reacting to the functional decline, to predicting risk early and proactively taking steps to prevent deterioration.

Current clinical strategies include the repeated use of clinical scales such as the short physical performance battery (SPPB) score, Modified Barthel Index and the Timed Up and Go test[6,7]. Due to an absence of specific biomarkers, the role of serially tracking surrogate biomarkers of inflammation (interleukin-6, C-reactive protein), muscle degradation (CK-MB, myostatin), and frailty or malnutrition (albumin, prealbumin), have also been explored[6].

With most of our understanding of HAFD, it becomes evident that the problem is grave and our understanding sparse. Thus, in order to improve outcomes, validated clinical prediction models to identify those vulnerable to HAFD forms a research priority.

Machine learning (ML) has a distinct advantage over clinical judgement and scores derived from conventional regression models, since it has the ability to detect subtle non-linear interactions, such as disproportionate risk escalation at advanced age in the presence of frailty or prolonged immobilization. ML models are designed to look for patterns across multitudes of variables. A forecast based on ML models benefit by transforming routinely collected clinical data into early, actionable risk signals[5]. ML models for HAFD can help predict which patients are likely to experience meaningful deterioration in mobility or ADL performance between admission and discharge.

Electronic medical records (EMRs) contain a rich stream of signals-vital sign trends, laboratory variations, medication exposure, nursing notes, activity levels, mobility documentation, length of bed rest, therapy referrals, delirium flags, and even patterns of missed meals or sleep disturbances. Analysis of such vast and varied data maybe beyond the human capacity of a clinician. ML-based HAFD prediction models can utilise this vast data by integrating pre-operative variables, namely baseline demographics, any geriatric syndromes, comorbidities, with perioperative inflammatory markers, nursing assessments, and patterns within electronic medical records, enabling early ongoing identification, risk stratification, and targeted prevention pathways. This may also help to individualize risk prediction score to different surgeries. Individualised prescriptions for prehabilitation including physiotherapy and active exercises, nutritional buildup, comorbidities optimisation and counselling can then be generated for such geriatric patients. Postsurgical rehabilitation pathways, right up to discharge planning and at home follow-ups, can be similarly individualised. In combination with enhanced recovery protocols for major surgery, they may help ensure a meaningful recovery for our elderly patients[8]. This forms a complete roadmap from prediction to prevention for HAFD. Such an effective, dynamic and updated risk estimate model can support decision-making at the bedside by helping clinicians answer some crucial questions: Which patient needs intervention today so that the patient may return home in the same or better condition than in which he arrived at the hospital?

Hiramatsu et al[7] have addressed the question on preoperative prediction of HAFD in geriatric patients undergoing cardiovascular surgery. They have retrospectively analysed the data of 144 patients out of which 41 (28.5%) patients experienced HAFD. They have identified seven preoperative clinical predictors (age, gender, body mass index, preoperative walking speed, SPPB orthostasis, SPPB balance, and surgery type) using LASSO regression. They have further divided their data (n = 144) into 75% training (n = 108) and 25% testing (n = 36) sets and constructed predictive models comparing seven ML algorithms, to identify the ML algorithm that works best for the given dataset. They have used both area under the curve (AUC) as well as SHapley Additive exPlanations (SHAP) values to interpret the results and found extreme gradient boosting (XGBoost) to be the best fit. They report that female gender and presence of a slow preoperative walking speed were significant factors influencing the model’s predictions.

Although ML-based tools appear promising, certain concerns need to be addressed prior to clinical adoption[9]. The quality of data has a high impact on the model performance. Small retrospective datasets, prone to missing, incomplete or inconsistently coded data, can introduce bias and reduce reliability[9]. Numerous clinical predictors and varied definitions or measuring scales of the outcome variable limit the generalisability and real-world reliability. LASSO regression narrows the selection of clinical predictors or features, a term used in ML language, by adding penalty to shrink the coefficients, so that weak predictors get automatically eliminated. When many colinear clinical predictors are present, especially in smaller datasets, LASSO is thus useful for better selection of screening variables which can then be utilised in ML algorithms including multiple logistic regression[5]. The assessment of the functional decline may also be influenced by staffing, documentation practices or language barriers in retrospective datasets. ML models inadvertently learn these spurious associations driven by confounding or institutional practices[9].

When developing and evaluating new ML algorithms, a minimum of two datasets (training models and testing models) are required[5,9]. The testing dataset should be different from the training dataset and representative of the wider population. The small sample size of the testing dataset along with training and testing sets both being a subset of the same population can reduce the generalisability of the model[6,9]. Overfitting with ML remains a key concern[9]. Multiple validations in separate larger datasets, tested in different populations, spread across multiple regions, need to be performed to develop robust models[9].

Several ML models exist and their performance depends on the type of data fed into them. XGBoost, CatBoost and AdaBoost are more complex models, useful to detect interactions between complex non-linear variables, which provide better testing metrics (AUC, calibration, sensitivity, specificity)[5]. But their major drawback is their limited interpretability and understanding of the process. Clinicians are hesitant to act on a “black-box” prediction unless they understand why the patient is flagged as high risk[5,9]. The use of explainable artificial intelligence such as SHAP, is widely acknowledged as a crucial element to navigate this complex landscape. SHAP is a feature attribution technology, which provides a numeric assessment of each feature’s (input variable’s) contribution to the model development, thus rendering the “black-box” prediction label moot[10]. Reporting of ML-based prediction models using TRIPOD-AI statement promotes transparent and complete reporting, can further build clinician trust and improve widespread acceptability[11].

Effective clinical adoption requires local calibration, validation by means of multi-centric studies, embedding the tool within EMR workflows, ensuring clinician engagement while listing clear recommended actions, and demonstrating clinically meaningful functional outcomes. Finally, continuous monitoring and recalibration over time are essential to ensure enduring prediction models.

CONCLUSION

In conclusion, use of ML-based algorithms for prediction of HAFD is practical and feasible. Development of prediction models can pave the way for building models that guide prevention strategies. With the help of ML, which enables integration of humungous amount of perioperative data, such prediction-prevention models can mitigate patient risk and personalise care. If developed responsibly, they can support individualized perioperative care pathways, aligning with what patients truly value: Not merely discharge or survival, but a return to meaningful quality of life.