Application of machine learning approaches in predicting clinical outcomes in older adults – a systematic review and meta-analysis.

doi:10.21203/rs.3.rs-2095524/v1

Download PDF

Research Article

Application of machine learning approaches in predicting clinical outcomes in older adults – a systematic review and meta-analysis.

https://doi.org/10.21203/rs.3.rs-2095524/v1

This work is licensed under a CC BY 4.0 License

Journal Publication

published 14 Sep, 2023

Read the published version in BMC Geriatrics →

You are reading this latest preprint version

Background

Machine learning-based prediction models have the potential to have a considerable positive impact on geriatric care.

Design:

Systematic review and meta-analyses.

Participants:

Older adults (≥ 65 years) in any setting.

Intervention:

Machine learning models for predicting clinical outcomes in older adults were evaluated. A meta-analysis was conducted where the predictive models were compared based on their performance in predicting mortality.

Outcome measures:

Studies were grouped by the clinical outcome, and the models were compared based on the area under the receiver operating characteristic curve metric.

Results

29 studies that satisfied the systematic review criteria were appraised and six studies predicting a mortality outcome were included in the meta-analyses. We could only pool studies by mortality as there were inconsistent definitions and sparse data to pool studies for other clinical outcomes. The area under the receiver operating characteristic curve from six studies included in the meta-analysis yielded a summary estimate of 0.82 (95%CI: 0.76–0.87), signifying good discriminatory power in predicting mortality.

Conclusion

The meta-analysis indicates that machine learning models can predict mortality. As electronic healthcare databases grow larger and more comprehensive, the available computational power increases and machine learning models become more sophisticated; they should be integrated into a larger research setting to predict various clinical outcomes.

Older adults aged ≥ 65 years are the highest healthcare consumers, accounting for the largest proportion of hospitalisations and in-hospital deaths(1). Moreover, this growing population exhibits the highest levels of multimorbidity(2). There is an immediate need to develop accurate machine learning algorithms for predicting unfavourable clinical outcomes in older adults. Given that clinical outcomes can be accurately predicted using machine learning, medical resources can be managed more specifically, prioritising those with a high risk of an unfavourable clinical outcome. Machine learning has already been applied in geriatric medicine for various purposes, including but not limited to cancer diagnosis(3), predicting falls(4), mortality(5), length-of-stay with chronic diseases at the point of admission(6), breast cancer risk(7), cancer prognosis(8) and response to chemotherapy in cancer patients(9).

While several studies discuss machine learning in the context of geriatric care, most of these approaches concern clinical diagnostics. There is limited literature on machine learning to predict key clinical outcomes that will directly inform improvements in geriatric care and prescribing policy. Machine learning can be applied to several areas, such as predicting frailty, and memory impairment, identifying patients at risk of heart failure and predicting adverse drug events. For example, successfully predicting the severity of frailty in an older patient allows clinicians to take proactive measures to minimise potential harm(10) or triage patients for more intensive treatment options. Likewise, predicting the risk factors for heart failure allows a clinician to make recommendations to decrease the likelihood of a poor clinical outcome. The scope of machine learning in geriatric care is broad, and with further advancement in methodology, hardware and medical database infrastructure, machine learning can have a significant positive effect on geriatric care.

Machine learning uses algorithms to analyse vast amounts of input data and make predictions based on satisfying predetermined criteria. However, traditional statistical approaches for predicting clinical outcomes are parsimonious and possess predetermined modelling assumptions, a drawback from which machine learning does not suffer. This paper aims to provide an overview of the literature on predicting clinical outcomes in older adults using machine learning. A particular focus was put on adults ≥ 65 years of age due to a lack of literature relating to this population. Applying machine learning to these populations allows clinicians to identify high-risk patients and distribute valuable medical resources accordingly, considerably improving clinical care in older people.

This systematic review was conducted after the study protocol was registered with PROSPERO (CRD42021295956).

Search Strategy

For this systematic review and meta-analysis, studies describing the use of machine learning models in predicting clinical outcomes in older adults (≥ 65 years) were assessed for predictive validity and reliability. The literature databases included in the study were PubMed, Embase, Web of Science core collection, Web of Science BIOSIS citation index, Scopus and ProQuest, using keywords in titles, abstracts, and index terms. This systematic review and meta-analysis used all relevant literature published in English up to the 30th of November 2021. All studies were uploaded to EndNote Version 9 for duplicate removal. After removing the duplicates, the remaining studies were uploaded to Covidence software for abstract screening, full-text review, data extraction and quality assessment. The complete search strategy can be viewed in Supplementary information Table 2.

Inclusion and Exclusion Criteria

Predefined criteria were established for the inclusion and exclusion of studies. Regarding the participants, only older adults (years of age ≥ 65) were included. Studies concerning adult-only populations, animal studies, and pre-term populations were excluded. Regarding the intervention, studies which applied machine learning-based clinical prediction algorithms in older adults were included. Any synonyms for machine learning, such as “deep learning” or “statistical learning,” were included. The application of machine learning to non-clinical settings, application of machine learning solely to images and/or signals, and studies that did not use machine learning were excluded. Regarding comparators, studies concerning machine learning vs other machine learning methods, machine learning vs traditional statistical, clinical prediction tools, and machine learning vs the unaided clinician were included. Studies utilising only traditional statistical prediction tools or unaided clinician performance alone were excluded. Regarding the type of study, cohort studies (retrospective and prospective), cross-sectional studies and grey literature about the implementation of machine learning in clinical prediction tools in the geriatric setting were included. Narrative reviews, letters, abstracts only, corrigendum, no full-text available studies, and practice guidelines were excluded.

The complete methods section, including the study selection process, data extraction and model performance, quality assessment of included studies, statistical analysis and deviations from the PROSPERO protocol, can be seen in Supplementary Information Appendix 1: Methods (cont).

Study Identification

The search identified 6882 studies. After removing duplicates, 4033 studies underwent title and abstract screening. From the title and abstract screen, 1347 studies progressed to full-text review, and 2686 were considered irrelevant. From the full-text review, 29 studies progressed to data extraction, and 1318 were excluded. The reasons for exclusion were the wrong patient population (n = 1087), wrong intervention (n = 95), wrong study design (n = 66), wrong outcomes (n = 39), conference abstract (n = 18), review article (n = 10), letter to the editor (n = 2), study protocol (n = 1). Figure 1 shows the study selection flowchart.

Study Characteristics

The characteristics of the studies included in this review are summarised in Table 1 (page 8). Most studies (n = 17) originated from the USA, with the remaining studies originating from various EU countries as well as Taiwan, Japan, China and Iceland. All studies were published post-1994, with the majority of studies published in the last 5 years (n = 21). The data was mainly acquired from electronic healthcare records (EHR) (n = 17). The majority of studies were of a retrospective design. The most common study design was case-control studies (n = 12), and the least common study design was a descriptive study, longitudinal prospective study and retrospective case-control study (n = 1 each). There was a large variation in sample sizes, ranging from 15 to 1,214,489 patients. All studies concerned exclusively patients aged ≥ 65 years.

Outcomes

The primary outcome varied across studies, with most (n = 14) predicting a mortality outcome. Other commonly reported outcomes included clinical care admission/readmission (n = 3), delirium (n = 2) and survival (n = 2). Due to a limited number of studies, a meta-analysis was only performed for the mortality outcome.

Quality Assessment

2 studies failed to meet the QUADAS-2 ≥ 11/14 score threshold for inclusion in the meta-analysis, scoring 8/14. 23 studies scored 14/14, 1 study scored 13/14, 3 studies scored 12/14, and 2 scored 11/14. Supplementary Information Table 4 shows the performance of each study in the QUADAS-2 assessment.

ML Models

A wide variety of machine learning approaches were used across the studies. The most common machine learning approaches utilised included logistic regression (n = 10), Random Forest (RF) (n = 7), decision trees (n = 6), Artificial Neural Network (ANN) (n = 4), XGBoost (n = 3).

ML Model Performance and Evaluation

All studies utilised a classical approach for appraising the performance of the model. Most commonly, Area Under the Receiver Operator Characteristic Curve (AUC-ROC) was utilised in 22 studies. Other commonly used methods for appraising performance included sensitivity (n = 9), specificity (n = 9), accuracy (n = 6), Root Mean Square Error (RMSE) (n = 2), Mean Square Error (MSE) (n = 1), and F1 values (n = 1).

Meta-Analysis

Six studies were included in the meta-analysis. Studies were grouped by the outcome (mortality) and compared based on the AUC-ROC. The summative results can be seen in Fig. 2. Typically AUC-ROC values between 1–0.9 are considered ‘excellent’, 0.9 − 0.8 ‘good’, 0.8 − 0.7 ‘fair’ 0.7 − 0.6 ‘poor’ and 0.6 − 0.5 ‘failed’. The highest AUC-ROC were recorded by Gomes et al., 2021, where ANN, Support Vector Machine (SVM) and RF models were applied to predict all-cause intrahospital mortality after Transcatheter Aortic Valve Implantation (TAVI). The most effective model was RF with an AUC-ROC of 0.97 [95%CI: 0.95–0.98], followed by ANN (AUC-ROC of 0.96 [95%CI: 0.94–0.97]) and SVM (AUC-ROC of 0.94 [95%CI: 0.91–0.96]). The models presented by Diaz-Ramirez et al., 2021 showed the lowest AUC-ROC readings from the meta-analysis, ranging from an AUC-ROC of 0.66 for predicting mortality using the intersection method to 0.72 for predicting mortality using the full method model. Several studies, such as Duarte et al., 2015, Sanclaro et al., 2012 and Pilotto et al., 2010 showed a large variety of AUC-ROC results. The complete key for Fig. 2 can be seen in Supplementary Information Appendix 2.

Method	t-statistic	df	p
Egger (unweighted)	1.929	17	0.071
Egger (multiplicative overdispersion)	0.544	17	0.594

Funnel plots for the studies included in the meta-analysis show the variability of individual studies (standard error) versus the mean effect size, as seen in Fig. 3. In both the unweighted and multiplicative overdispersion plots, there is good symmetry with an equal dispersion of studies below and above the mean. Moreover, some studies fall outside the funnel in both the unweighted and multiplicative overdispersion plots, indicating publication bias. It must be noted that Egger’s test(11) is less precise in detecting bias, given the small number of studies.

Our review identified 29 studies which utilised a machine learning approach to predict a clinical outcome in adults aged ≥ 65 years. The meta-analysis consisted of six studies utilising different machine learning models to predict mortality. Supported by an AUC-ROC summary estimate of 0.82 (95%CI: 0.76–0.87), we found that machine learning models perform better than traditional statistical approaches regarding discriminatory power. This finding can be coupled with the fact that as better data and more sophisticated hardware associated with machine learning become available, so will the predictive power of these models. We concluded that there is scope for introducing predictive machine learning models into routine care to predict clinical outcomes in older adults.

Impact on Clinical Care of Older Adults

Machine learning and its improving ability to analyse vast amounts of patient data and make accurate predictions will soon become another tool in the armamentarium of an experienced physician, refining the gold standard in geriatric care. Machine learning has already shown promising results in predicting outcomes such as morbidity, hospitalisation, and length-of-stay; machine learning models should promptly be integrated into the clinical workflow. Perhaps the most impactful use of this technology would be identifying patients at high risk of falls. Around a third of people aged > 65 years experience an accidental fall at least once per year, with 20% of cases experiencing injury and 5% requiring hospitalisation(12). If patients were identified as ‘at risk’ early on, measures could be put in place to minimise the fall risk. Another use of machine learning could be tackling the issue of adverse drug reactions in older polypharmacy patients. In many instances, adverse drug reactions experienced by older adults go unreported or are attributed to a confounding factor. If adverse drug reactions could be predicted, measures such as changes in prescribing policy could be taken to minimise the potential harm to the patient. Finally, machine learning could be coupled with wearable devices such as smartwatches, acting as an early warning system for patients to seek medical assistance.

There are several key drawbacks to using machine learning which must be considered. Ideally, machine learning models could be applied to any EHR/clinical database and make live predictions as new data is introduced. However, since many EHR are fundamentally different in data structures, this will prove challenging. First, EHRs and clinical databases must be standardised to take full advantage of these approaches. Secondly, some machine learning algorithms require much larger amounts of high-quality data to make accurate predictions than traditional statistical approaches. Finally, most machine learning algorithms need to be optimised and the optimal model tuning would be user dependent which may introduce additional challenge in assessing each model’s performance and risk of bias.

Predictive Model Methods

The three best performing models in the meta-analysis were part of a study by Gomes et al., 2021 in which an RF (AUC-ROC 0.97), an SVM (AUC-ROC 0.94) and an ANN (AUC-ROC 0.96) were applied to predict all-cause intrahospital mortality after TAVI in a cohort of 451 patients. These AUC-ROC values were obtained using all available features (n = 83). The same models performed slightly worse using baseline features (n = 50); AUC-ROC 0.81, 0.83, 0.72, respectively. All three models significantly outperformed traditional logistic risk scores, the Society of Thoracic Surgeons score and the Society of Thoracic Surgeons-American College of Cardiology Transcatheter Valve Therapy Registry score, which performed at AUC-ROC 0.64 and 0.65, respectively. It can be concluded that a higher number of input variables corresponds to a higher AUC-ROC. The small sample size of patients limits this study’s validity and generalisability. In a study by Morris et al., 2020 a two-tiered scoring system was developed to predict all-cause intrahospital mortality. The quick Elderly Mortality After Trauma (qEMAT) is for use upon initial presentation, and the full Elderly Mortality After Trauma (fEMAT) is for after radiologic evaluation. In the study, patients were split into derivation/validation cohorts, with no overlap between patients. The qEMAT reached an AUC-ROC of 0.80 in both cohorts, and the fEMAT reached an AUC-ROC of 0.85 in both cohorts. The qEMAT was also tested in a separate dataset, the Michigan Trauma Quality Improvement Program, which achieved AUC-ROC scores of 0.82, outperforming its result in the derivation and validation cohorts. Pilotto et al., 2010 applied Multidimensional Prognostic Index (MPI) to predict 1-month mortality in 376 patients with a diagnosis of heart failure split into cohorts by sex. Traditional regression model scores were also calculated. However, MPI performed best, reaching an AUC-ROC of 0.80 in women and 0.83 in men. In a study by Sancarlo et al., 2012 MPI was used to predict 1-month, 6-month and 12-month all-cause mortality. MPI showed a significant high discriminatory power with AUC-ROC of 0.819, 0.799 and 0.770, respectively. Duarte et al., 2015 describe the Patient-Reported Outcome Mortality Prediction Tool (PROMPT) in predicting 6-month mortality. Patients were split into derivation/validation cohorts. PROMPT achieved similar AUC-ROC readings in both cohorts, 0.75 and 0.73, respectively. Duarte concluded that PROMPT demonstrates good discrimination but poor calibration in independent heterogeneous datasets. Finally, Diaz-Ramirez et al., 2021 used the BI (baBIC) to predict mortality. A ‘Union Method’ consisted of a subset of common predictors obtained from the union of optimal models for each outcome. A subset obtained from the intersection of optimal models for each outcome made up the ‘Intersection method’, and a model with no variable selection made up the ‘Full Method’. The best performance was recorded from the ‘Full Method’, AUC-ROC 0.72, although the other approaches showed similar results (AUC-ROC 0.71 for all other methods except the Intersection method, AUC-ROC 0.66).

Limitations and Strengths

Our approach has several limitations. First, the search was limited to English, meaning relevant literature in other languages was not considered. Second, the review could have missed some studies due to unusual terminology. To mitigate this, significant time was spent identifying all synonyms of keywords and working with an experienced subject librarian when designing the search strategy. Third, due to limited literature in the field, only six studies were included in the meta-analysis. Fourth, there is little heterogeneity between the patients from the studies included in the meta-analysis, and several patients had co-morbidities such as heart failure. Fifth, we could only pool studies by one outcome (mortality) as there were inconsistent definitions and sparse data to pool studies for other clinical outcomes. Sixth, inconsistent reporting of the model development process and inconsistent reporting of model performance meant that some studies appropriate for meta-analysis were not included. For instance, not all performance metrics can be converted to AUC-ROC, especially without specific data on the number of subjects in the derivation/validation cohorts. Finally, the predictions were performed on retrospective data, mostly from electronic health records and validating these models on prospectively collected data from other sources is critical for reproducibility. Our study’s strengths include following PRISMA guidelines, having our protocol approved and registered by PROSPERO, using Covidence software (increased rigour, accurate documentation), using a highly appropriate quality of study tool (QUADAS-2) and utilising multiple literature databases.

Given the vast amounts of clinical data collected from patients, there is considerable potential in utilising machine learning in predicting clinical outcomes in older adults. This review showed that machine learning models exhibit significantly higher discriminatory power than traditional statistical approaches. It should also be considered that as datasets become larger and machine learning models become more sophisticated, their performance will increase further. As this field of research develops, we believe it is important that specific attention is dedicated to a comprehensive model performance analysis for each model. More large-scale studies are needed to show that machine learning can predict clinical outcomes in older adults underrepresented in our findings, such as the prediction of risk factors for dementia, identifying risk factors for multimorbidity, and screening subgroups of older adults vulnerable to falls and fractures.

Table 1

*Executive summary of the table of study characteristics. The full table of study characteristics can be found in* Supplementary Information Table 1.
Authors; Year; Country	Sample Size	Primary Outcome	ML Approach	ML Assessment Metric
Bowen et al; 2021; USA(13)	15	Pain perception.	Multilevel mixed modelling.	AUC-ROC.
Chen et al; 2020; China(14)	964	Survival in patients with early stage uterine papillary serous carcinoma.	Proportional subdistribution hazards regression.	The c-index.
Chung et al; 2020; Taiwan(15)	31	mRS, mortality rate at 3 months.	Two ANN models.	AUC-ROC, correlation efficiency, MSE.
Considine et al; 2019; Australia(16)	1,717	Transfer characteristics and outcomes, serious adverse events and mortality.	Multivariable logistic regression.	AUC-ROC, RMSE, mean absolute errors, AIC, BIC, Hosmer-Lemeshow statistics.
Das et al; 2003; USA(17)	190	Intervention for control of haemorrhage, recurrent bleeding, and death.	ANN, multiple-logistic regression, BLEED classification system.	Accuracy, sensitivity, specificity, PPV, NPV, likelihood ratios for positive and negative tests, discriminant power. McNemar’s test, AUC-ROC.
Diaz-Ramirez et al; 2021; USA(18)	5,531	Multiple outcomes including admission and mortality.	baBIC.	Monte Carlo standard error, c-statistic.
Duarte et al; 2015; USA(19)	467	6-month mortality.	PROMPT.	AUC-ROC, Hosmer-Lemeshow statistic, sensitivity, specificity, PPV, NPV, positive and negative likelihood ratios.
Falsetti et al; 2021; Italy(20)	1,326	Therapeutic failure.	XGBoost.	AUC-ROC.
Ford et al; 2021; UK(21)	46,713	Dementia diagnosis code.	Logistic regression, naive Bayes, RF.	AUC-ROC, PPV.
Ghffar et al; 2020; USA(22)	354	In-hospital cardiovascular mortality, In-hospital all-cause mortality, 30-day cardiovascular mortality, 30-day all-cause mortality.	Decision trees, ensembles, logistic regressions, deep nets.	Specificity, sensitivity, false-negative rate, and AUC-ROC.
Ghotra et al; 2020; USA(23)	1,177	1-year mortality.	Backward propagation neural network, augmented multivariable regression.	Hosmer-Lemeshow goodness-of-fit test, Akaike and Schwarz BIC, and AUC-ROC.
Gomes et al; 2021; Germany(24)	451	All-cause intrahospital mortality.	ANN, SVM, RF.	AUC-ROC.
Han et al; 2012; USA(25)	21,870	6-month mortality.	Logistic regression.	k-fold cross-validation, AUC-ROC, specificity, PPV, NPV, positive and negative likelihood ratios.
Hyer et al; 2020; USA(26)	682,114	Various outcomes including 30/90 day readmission and 30/90 day mortality.	Decision trees.	Accuracy.
Ko et al; 2014; USA(27)	14,844	Colonoscopy indication.	Decision trees, linear discriminant analysis.	Accuracy, sensitivity, specificity.
Morris et al; 2020; USA(28)	1,214,489	In-hospital mortality.	qEMAT, fEMAT.	Calibration and AUC-ROC.
Ocagli et al; 2021; Italy(29)	78	4AT delirium score.	RF.	RMSE.
Parenica et al; 2012; Czech Republic(30)	45	Adverse clinical outcomes after TAVI and SAVR.	EuroSCORE	AUC-ROC.
Pilotto et al; 2010; Italy(31)	376	1-month mortality.	MPI combined with the linear regression Acute Decompensated Heart Failure National Registry model.	AUC-ROC.
Pompei et al; 1994; USA(32)	432	Delirium.	Logistic regression.	AUC-ROC.
Rossi et al; 2021; Italy(33)	1794	Myeloid neoplasm development.	Multivariable cox analysis.	C-index, PPV, NPV, AUC-ROC.
Sancarlo et al; 2012; Italy(34)	654	All-cause mortality, 6/12-month all-cause mortality.	MPI	AUC-ROC.
Sax et al; 2021; USA(35)	26,189	Adverse events within 30 days of emergency department arrival.	Logistic regression, LASSO, decision tree, RF, and XGBoost.	AUC-ROC, sensitivity, specificity, negative and positive likelihood ratios, PPV, NPV, F1 values.
Shardell et al; 2021; USA(36)	16,388	Sex-specific serum 25-hydroxyvitamin D thresholds.	Weighted decision trees.	Sensitivity.
Suzuki et al; 2020; Japan(37)	504	180-day all-cause mortality.	Multiple logistic regression.	AUC-ROC.
Velagapudi et al; 2021; USA(37)	220	Binarised thrombolysis in cerebral infarction on first pass.	Logistic regression, RF, SVM, Naive Bayes, and XGBoost.	AUC-ROC, accuracy scores, Matthews correlation coefficient, Brier score, sensitivity, specificity.
Venturini et al; 2021; Italy(38)	176	Various clinical outcomes including mortality, intensive care unit admission and intensive care unit discharge.	A conditional inference RF model, a classic RF model, an ordinal forest model, decision trees model, a recursive partitioning regression tree model.	Accuracy, sensitivity, specificity.
Zarkowsky et al; 2021; USA(39)	25,376	Short-stay discharge appropriateness, survival at 30 days.	Multivariable statistical analyses, multilayer perceptron.	AUC-ROC, Harrell’s c-statistic.
Zhou et al; 2021; USA(40)	149	Heart failure with preserved ejection fraction risk.	GA-KPLS, LASSO, RF, ridge regression, SVM, and logistic regression.	AUC-ROC, sensitivity, specificity, accuracy, Youden index, G-means, and Matthews correlation coefficient, one-way ANOVA, Dunnett’s multiple-comparison test.

Table 1. Abbreviations: C-statistic – concordance statistic, PPV – Positive Predictive Value, NPV – Negative Predictive Value, ANOVA - Analysis of Variance, NR – Not reported, SAF - Standard Analytic Files, mRS - Modified Rankin Scale, BIC - Bayesian Information Criterion, CORI - Clinical Outcomes Research Initiative, MHOS - Medicare Health Outcomes Survey, GA-KPLS - Kernel partial least squares with the genetic algorithm, LASSO - Least Absolute Shrinkage and Selection Operator, SEER - Surveillance, Epidemiology, and End Results, HRS - Health and Retirement Study, SAVR - Surgical Aortic Valve Replacement.

Declarations

Ethics approval and consent to participate

Not applicable.

Consent for publication

Not applicable.

Declarations

Competing interests

The authors declare no competing interests.

Funding

URSA (University Research Studentship Award) at the University of Bath.

Authors’ contributions

Search strategy design; RO. Abstract Screen; RO. Full-text screen; RO, SR, PN. Quality assessment and risk of bias assessment; RO. Manuscript writing; RO under the guidance of SR and PN. Final editing; RO under the guidance of SR and PN. All authors critically reviewed the content of the literature review and approved its final version.

Acknowledgements

We thank Tamsin Riley, the subject librarian for Pharmacy and Pharmacology at the University of Bath, for assisting in designing the literature search strategy.

Dixon T, Shaw M, Frankel S, Ebrahim S. Hospital admissions, age, and death: retrospective cohort study. BMJ. 2004;328(7451):1288.
Kingston A, Robinson L, Booth H, Knapp M, Jagger C, project M. Projections of multi-morbidity in the older population in England to 2035: estimates from the Population Ageing and Care Simulation (PACSim) model. Age Ageing. 2018;47(3):374 – 80.
Capper D, Jones DTW, Sill M, Hovestadt V, Schrimpf D, Sturm D, et al. DNA methylation-based classification of central nervous system tumours. Nature. 2018;555(7697):469–74.
Nait Aicha A, Englebienne G, van Schooten KS, Pijnappels M, Krose B. Deep Learning to Predict Falls in Older Adults Based on Daily-Life Trunk Accelerometry. Sensors (Basel). 2018;18(5).
Tedesco S, Andrulli M, Larsson MA, Kelly D, Alamaki A, Timmons S, et al. Comparison of Machine Learning Techniques for Mortality Prediction in a Prospective Cohort of Older Adults. Int J Environ Res Public Health. 2021;18(23).
Hu Z, Qiu H, Wang L, Shen M. Network analytics and machine learning for predicting length of stay in elderly patients with chronic diseases at point of admission. BMC Med Inform Decis Mak. 2022;22(1):62.
Ming C, Viassolo V, Probst-Hensch N, Dinov ID, Chappuis PO, Katapodi MC. Machine learning-based lifetime breast cancer risk reclassification compared with the BOADICEA model: impact on screening recommendations. Br J Cancer. 2020;123(5):860–7.
Yokoyama S, Hamada T, Higashi M, Matsuo K, Maemura K, Kurahara H, et al. Predicted Prognosis of Patients with Pancreatic Cancer by Machine Learning. Clin Cancer Res. 2020;26(10):2411–21.
Clayton EA, Pujol TA, McDonald JF, Qiu P. Leveraging TCGA gene expression data to build predictive models for cancer drug response. BMC Bioinformatics. 2020;21(Suppl 14):364.
Aponte-Hao S, Wong ST, Thandi M, Ronksley P, McBrien K, Lee J, et al. Machine learning for identification of frailty in Canadian primary care practices. Int J Popul Data Sci. 2021;6(1):1650.
Matthias Egger GDS, Schneider M. Christoph Minder. Bias in meta-analysis detected by a simple geographical test. BMJ. 1997;315:629–34.
Abell JG, Lassale C, Batty GD, Zaninotto P. Risk Factors for Hospital Admission After a Fall: A Prospective Cohort Study of Community-Dwelling Older People. J Gerontol A Biol Sci Med Sci. 2021;76(4):666–74.
Bowen ME, Ji X, Griffioen MA. Poor Sleep Predicts Increased Pain Perception Among Adults With Mild Cognitive Impairment. Nurs Res. 2021;70(4):310–6.
Chen L, Liu X, Li M, Wang S, Zhou H, Liu L, et al. A novel model to predict cancer-specific survival in patients with early-stage uterine papillary serous carcinoma (UPSC). Cancer Med. 2020;9(3):988–98.
Chen-Chih Chung Y-CC, Hong C-T, Hu C-J, Chan Han-HwaHLung. Hung-Wen Chiu. Artificial neural network-based analysis of the safety and efficacy of thrombolysis for ischemic stroke in older adults in Taiwan. Neurology Asia. 2020.
Considine J, Street M, Bucknall T, Rawson H, Hutchison AF, Dunning T, et al. Characteristics and outcomes of emergency interhospital transfers from subacute to acute care for clinical deterioration. Int J Qual Health Care. 2019;31(2):117–24.
Das A, Ben-Menachem T, Cooper GS, Chak A, Sivak MV, Gonet JA, et al. Prediction of outcome in acute lower-gastrointestinal haemorrhage based on an artificial neural network: internal and external validation of a predictive model. The Lancet. 2003;362(9392):1261–6.
Diaz-Ramirez LG, Lee SJ, Smith AK, Gan S, Boscardin WJ. A Novel Method for Identifying a Parsimonious and Accurate Predictive Model for Multiple Clinical Outcomes. Comput Methods Programs Biomed. 2021;204:106073.
Duarte CW, Black AW, Murray K, Haskins AE, Lucas L, Hallen S, et al. Validation of the Patient-Reported Outcome Mortality Prediction Tool (PROMPT). J Pain Symptom Manage. 2015;50(2):241–7. e6.
Falsetti L, Rucco M, Proietti M, Viticchi G, Zaccone V, Scarponi M, et al. Risk prediction of clinical adverse outcomes with machine learning in a cohort of critically ill patients with atrial fibrillation. Sci Rep. 2021;11(1):18925.
Ford E, Sheppard J, Oliver S, Rooney P, Banerjee S, Cassell JA. Automated detection of patients with dementia whose symptoms have been identified in primary care but have no formal diagnosis: a retrospective case-control study using electronic primary care records. BMJ Open. 2021;11(1):e039248.
Abdul Ghffar Y, Osman M, Shrestha S, Shaukat F, Kagiyama N, Alkhouli M, et al. Usefulness of Semisupervised Machine-Learning-Based Phenogrouping to Improve Risk Assessment for Patients Undergoing Transcatheter Aortic Valve Implantation. Am J Cardiol. 2020;136:122–30.
Ghotra AS, Monlezun DJ, Boone D, Jacob R, Poosti K, Loghin C, et al. Outcomes of Patients Undergoing Transcatheter Aortic Valve Implantation With Incidentally Discovered Masses on Computed Tomography. Am J Cardiol. 2020;132:114–8.
Gomes B, Pilz M, Reich C, Leuschner F, Konstandin M, Katus HA, et al. Machine learning-based risk prediction of intrahospital clinical outcomes in patients undergoing TAVI. Clin Res Cardiol. 2021;110(3):343–56.
Han PK, Lee M, Reeve BB, Mariotto AB, Wang Z, Hays RD, et al. Development of a prognostic model for six-month mortality in older adults with declining health. J Pain Symptom Manage. 2012;43(3):527–39.
Hyer JM, Paredes AZ, White S, Ejaz A, Pawlik TM. Assessment of utilization efficiency using machine learning techniques: A study of heterogeneity in preoperative healthcare utilization among super-utilizers. Am J Surg. 2020;220(3):714–20.
Ko CW, Dominitz JA, Neradilek M, Polissar N, Green P, Kreuter W, et al. Determination of colonoscopy indication from administrative claims data. Med Care. 2014;52(4):e21-9.
Morris RS, Milia D, Glover J, Napolitano LM, Chen B, Lindemann E, et al. Predictors of elderly mortality after trauma: A novel outcome score. J Trauma Acute Care Surg. 2020;88(3):416–24.
Ocagli H, Bottigliengo D, Lorenzoni G, Azzolina D, Acar AS, Sorgato S, et al. A Machine Learning Approach for Investigating Delirium as a Multifactorial Syndrome. Int J Environ Res Public Health. 2021;18(13).
Parenica J, Nemec P, Tomandl J, Ondrasek J, Pavkova-Goldbergova M, Tretina M, et al. Prognostic utility of biomarkers in predicting of one-year outcomes in patients with aortic stenosis treated with transcatheter or surgical aortic valve implantation. PLoS ONE. 2012;7(12):e48851.
Pilotto A, Addante F, Franceschi M, Leandro G, Rengo G, D'Ambrosio P, et al. Multidimensional Prognostic Index based on a comprehensive geriatric assessment predicts short-term mortality in older patients with heart failure. Circ Heart Fail. 2010;3(1):14–20.
Peter Pompei MF, Mark A, Rudberg SK, Inouye. Victoria Braund, and Christine K. Cassel. Delirium in Hospitalized Older Persons: Outcomes and Predictors. Journal of the American Geriatrics Society; 1994.
Marianna Rossi MM, 2 Matteo Zampini,1 Mauro Tettamanti,3 Emma Riva,3 Erica Travaglino,1 Matteo Bersanelli,4 Sara Mandelli,3 Alessia Antonella Galbussera,3 Ettore Mosca,5 Elena Saba,1 Chiara Chiereghin,1 Nicla Manes,1 Chiara Milanesi,1 Marta Ubezio,1 Lucio Morabito,1 Clelia Peano,1,6 Giulia Sold a,1,4 Rosanna Asselta,1,4 Stefano Duga,1,4 Carlo Selmi,1,4 Maria De Santis,1 Karolina Malik,4 Giulia Maggioni,1,4 Marilena Bicchieri,1 Alessia Campagna,1 Cristina A. Tentori,1,4 Antonio Russo,1,4 Efrem Civilini,1,4 Paola Allavena,1 Rocco Piazza,7 Giovanni Corrao,8 Claudia Sala,9,10 Alberto Termanini,1 Laura Giordano,1 Paolo Detoma,11 Aurelio Malabaila,11 Luca Sala,12 Stefano Rosso,13 Roberto Zanetti,13 Claudia Saitta,1,7 ElenaB Riva,1,7 Gianluigi Condorelli,1,4 Francesco Passamonti,14 Armando Santoro,1,4 Francesc Sole,15 Uwe Platzbecker,16 Pierre Fenaux,17 Niccol o Bolli,18,19 Gastone Castellani,9,10 Wolfgang Kern,2 George S. Vassiliou,20 Torsten Haferlach,2 Ugo Lucca,3 and Matteo G. Della Porta1,3,4. Clinical relevance of clonal hematopoiesis in persons aged OVER OR EQUAL TO 80 years. Blood. 2021.
Sancarlo D, Pilotto A, Panza F, Copetti M, Longo MG, D'Ambrosio P, et al. A Multidimensional Prognostic Index (MPI) based on a comprehensive geriatric assessment predicts short- and long-term all-cause mortality in older hospitalized patients with transient ischemic attack. J Neurol. 2012;259(4):670–8.
Sax DR, Mark DG, Huang J, Sofrygin O, Rana JS, Collins SP, et al. Use of Machine Learning to Develop a Risk-Stratification Tool for Emergency Department Patients With Acute Heart Failure. Ann Emerg Med. 2021;77(2):237–48.
Shardell M, Cappola AR, Guralnik JM, Hicks GE, Kritchevsky SB, Simonsick EM, et al. Sex-specific 25-hydroxyvitamin D threshold concentrations for functional outcomes in older adults: PRoject on Optimal VItamin D in Older adults (PROVIDO). Am J Clin Nutr. 2021;114(1):16–28.
Suzuki S, Motoki H, Kanzaki Y, Maruyama T, Hashizume N, Kozuka A, et al. A Predictive Model for 6-Month Mortality in Elderly Patients with Heart Failure. Int Heart J. 2020;61(2):325–31.
Venturini S, Orso D, Cugini F, Crapis M, Fossati S, Callegari A, et al. Classification and analysis of outcome predictors in non-critically ill COVID-19 patients. Intern Med J. 2021;51(4):506–14.
Zarkowsky DS, Nejim B, Hubara I, Hicks CW, Goodney PP, Malas MB. Deep Learning and Multivariable Models Select EVAR Patients for Short-Stay Discharge. Vasc Endovascular Surg. 2021;55(1):18–25.
Zhou L, Guo Z, Wang B, Wu Y, Li Z, Yao H, et al. Risk Prediction in Patients With Heart Failure With Preserved Ejection Fraction Using Gene Expression Data and Machine Learning. Front Genet. 2021;12:652315.

Abbreviations

95%CI	95% Confidence Interval
ANN	Artificial Neural Network
ANOVA	Analysis of Variance
AUC	Area Under the Curve
AUC-ROC	Area Under the Receiver Operator Characteristic (ROC) Curve
baBIC	Best Average Bayesian Information Criterion
BIC	Bayesian Information Criterion
CORI	Clinical Outcomes Research Initiative
CPRD	Clinical Practice Research Datalink
EHR	Electronic Healthcare Record
fEMAT	full Elderly Mortality After Trauma
GA-KPLS	Kernel partial least squares with the genetic algorithm
HRS	Health and Retirement Study
LASSO	Least Absolute Shrinkage and Selection Operator
MHOS	Medicare Health Outcomes Survey
mRS	Modified Rankin Scale
MPI	Multidimensional Prognostic Index
MMSE	Mini Mental State Examination
MSE	Mean Square Error
NPV	Negative Predictive Value
NR	Not Reported
PPV	Positive Predictive Value
PROMPT	Patient-Reported Outcome Mortality Prediction Tool
qEMAT	quick Elderly Mortality After Trauma
RF	Random Forest
RMSE	Root Mean Square Error
SAVR	Surgical Aortic Valve Replacement
SAF	Standard Analytic Files
SEER	Surveillance, Epidemiology, and End Results
SVM	Support Vector Machine
TAVI	Transcatheter Aortic Valve Implantation

No competing interests reported.

Download PDF

Journal Publication

published 14 Sep, 2023

Read the published version in BMC Geriatrics →

Editorial decision: Major revision
15 Mar, 2023
Reviews received at journal
10 Mar, 2023
Reviewers agreed at journal
09 Mar, 2023
Reviewers invited by journal
09 Mar, 2023
Editor assigned by journal
08 Mar, 2023
Editor invited by journal
17 Nov, 2022
Submission checks completed at journal
17 Nov, 2022
First submitted to journal
23 Sep, 2022

You are reading this latest preprint version

Application of machine learning approaches in predicting clinical outcomes in older adults – a systematic review and meta-analysis.

Status:

Journal Publication

Version 1

Abstract

Background

Design:

Participants:

Intervention:

Outcome measures:

Results

Conclusion

Background

Methods

Search Strategy

Inclusion and Exclusion Criteria

Results

Study Identification

Study Characteristics

Outcomes

Quality Assessment

ML Models

ML Model Performance and Evaluation

Meta-Analysis

Discussion

Impact on Clinical Care of Older Adults

Predictive Model Methods

Limitations and Strengths

Conclusion

Declarations

Declarations

Declarations

Funding

Authors’ contributions

Acknowledgements

References

Unsectioned Paragraphs

Unsectioned Tables

Additional Declarations

Supplementary Files

Status:

Journal Publication

Version 1