Machine learning in patient flow: a review

Patient Flow is a term used within healthcare services to refer to the way in which patients are moved through a healthcare facility. It involves the medical care, resources, and internal systems needed to get patients from admission to discharge while maintaining a standard of quality of care and satisfaction for the patient (Hall 2013). Many works have shown that patient flow can be predictable using machine learning techniques. These works aim to use these predictions to improve the flow of patients and resources in order to provide a faster and better service to patients.

The number of patient admissions to the hospital is arguably one of the most important aspects of patient flow. This determines the demand that is placed upon the hospital and therefore affects how patients can be treated. The importance of predicting patient admissions is reflected by the number of publications in this area. However, with little information on patients prior to their arrival it is also one of the most difficult areas of patient flow to create accurate predictions.

Boyle et al (2008, 2012) and Batal et al (2001) predict the number of emergency admissions using multiple regression. They frame the problem such that they forecast for daily admissions as well as weekly and monthly admissions. The use of regression is for interpretability of the predictions as well as the development of a simple model to improve the chance of being able to generalise to other hospitals. As mentioned previously, due to limited information on these patients prior to arrival, the authors use the days of the week and national holidays as features.

Whereas the aforementioned studies approach the problem as a static prediction (i.e. using information from a snapshot in time to make predictions), Tandberg and Qualls (1994), Au-Yeung et al (2009), and Schweigler et al (2009) treat the problem as a time-series. They use autoregressive models to account for the trajectory of the numbers of patients. This approach is more likely to be successful than a static approach due to the incorporation of data close to the event of interest. However, the benefit of a static approach (if the model is accurate) is that a prediction can be made at an early stage and action can be taken based on that prediction without needing to wait for the time-series to unfold. These time-series approaches also perform regressions to predict patient volumes in the coming days, weeks and months.

While the seasonal features such as weather and time of the year have been shown to be helpful with predicting patient numbers, they are not patient-specific and therefore are limited in their use for predicting when a patient will be admitted to hospital. As a result, LaMantia et al (2010) and Artetxe et al (2020) consider predicting patient readmissions to the ED instead of predicting any given admission. In doing so they are able to utilise the wealth of data already recorded by the hospitals on individual patients and identify markers that indicate high risk of readmission in an emergency. Hosseinzadeh et al (2013) use Naive Bayes and a decision tree in order to classify patients who are going to be readmitted to hospital using their health records as features. These works generally pose the readmission problem as predicting readmission within the next 30 d as this has the most impact on the health and welfare of the patient, as well as the scheduling of the hospital (Leppin et al 2014).

A problem that can arise due to these readmission predictions is that patients can be readmitted for various issues (for example a patient who was hospitalised for cardiac issues might need rehospitalisation for breaking their leg). Considering this type of readmission is not very useful for the hospital or the patient, as it is not indicative of an underlying condition and so the health records of the patient will not be useful for this prediction. To get around this issue, many authors have conditioned their prediction of admission on subsets of patients with certain underlying conditions. Shameer et al (2017) use a naive Bayes classifier to predict readmission and only considers a subset of patients with heart failure. They only consider a readmission to be valid if the patients are readmitted with heart failure within 30 d. Kalagara et al (2018) also condition their problem on a subset of patients who have had a neurosurgical procedure carried out and compare the performance of their model (trained used gradient boosted trees) using features available during the patients stay versus features that were obtained after the patients discharge. Naturally the model with access to features after the patient discharge performed better, however it is very difficult in most situations to obtain features post-discharge. Min et al (2019) carry out a similar study but consider patients suffering from COPD. They investigate various machine learning methods and find that gradient boosted trees offer the best prediction of readmission accuracy for their dataset. They also utilise recurrent neural networks in order to treat the problem as a time-series problem but the performance is significantly worse. In fact there are very few works that treat the prediction of readmission as a time-series due to the difficulty of obtaining data on patients post-discharge (Arora et al 2010).

The ultimate aim of all of the works mentioned in section 5.1 is to provide the hospital with an understanding of the volumes of patients that may be attending the ED. By forecasting this (and if the model is accurate) the hospitals may then plan the appropriate resource (including staff, tests and making equipment available) in order to be able to cope with the demand placed on them. For low numbers forecast, hospitals may also then reduce the required resource that is on standby which can lead to cost savings (Thungjaroenkul et al 2007).

Some authors however approach the problem from the scheduling perspective. This is different in that whereas predicting admissions makes the assumption that resource can be altered to meet demand, the scheduling approach does not. With this approach authors assume that there is fixed resource and how it is used can be optimised with varying patient numbers.

Rosemarin et al (2019) define the ED scheduling problem as needing to satisfy the following constraints: the schedule must minimise the risk of adverse consequences, minimise patient waiting time, minimise patient length-of-stay, minimise ED crowding and minimise interruption to caregivers. They use a mixture of health record data of the patients and data on the status the ED to reconstruct the state of the ED when the patients were there. They then use a mixed integer linear program to optimise these scenarios, maximising throughput while being constrained by the aforementioned constraints. They then train a deep learning architecture on this optimised data and use it as a ranking system to predict the optimal patient-caregiver pair in the ED.

Some authors prefer to allow the machine learning algorithms to discover the optimal policies instead of optimising the problem themselves to learn from. This is seen in Lee and Lee (2020a) where a deep Q network (a reinforcement learning algorithm) is used to learn the optimal policy of treating patients in the ED. In order to do this a simulation is made of the ED which will allow the agent to take exploratory moves essential for reinforcement learning. The state of the model is defined as the distribution of acuity (sum of patients at each acuity level) within the ED as well as the distribution of needed treatment type. The action of the agent is to rank the next patient that needs to be seen meaning it is also a patient priority-ranking system. Krämer et al (2019) also present a priority ranking system based on severity prediction, but go as far as predicting whether patient presentations to the ED should be treated as elective visits given their low severity. They do this using the primary diagnosis code of the patient, however there may be difficulties in expanding this tool to other hospitals given that many hospitals assign diagnosis codes after the patient is discharged from hospital and not at admission.

Yeh and Lin (2007) and Arisha and Abo-Hamad (2013) approach the scheduling problem slightly differently in that instead of ranking the priority patients in the ED, they instead aim to design the staffing schedules. They do this using genetic algorithms and allowing the staffing schedules to be updated and 'evolve' to a point where they are suitable for the demand placed on the ED. This approach makes the assumption that should the staffing level be predicted accurately, then there will be no need to prioritise patients in the ED as there will be enough staff (and resource) to process them.

Figure 2 shows the works that have been conducted so far on the prediction of admissions and scheduling in the ED. This is by no means an exhaustive summary but we aim to provide some structure to help other researchers understand what work has been conducted in the field of machine learning for patient flow through the ED. Table 1 further outlines the problems that have readily available datasets for prediction, and what models are popularly used to tackle the prediction problem in the literature. A lack of a readily available dataset for priority ranking is due to priority generally not being recorded in hospital EHRs. Readily available in this instance refers to existence in a typical hospital database and not that it is easily and openly accessible.

Zoom In Zoom Out Reset image size

Table 1. Popularity of different methods and data availability for each of these problems.

	ED-admission problem
	Volume pred.	Readmission pred.	Priority rank	Schedule
Labelled datatset readily available?	$\checkmark$	$\checkmark$	✗	$\checkmark$
Regression methods popular?	$\checkmark$	$\checkmark$	$\checkmark$	✗
Classification methods popular?	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$
Genetic methods popular?	✗	✗	✗	$\checkmark$

We have primarily focused on machine learning applied to emergency admissions as this is the larger body of research in the field. The stochastic nature of these admissions in terms of number and type of admission means that these are the most disruptive to patient flow in a hospital. Elective patients are generally planned for and so resource is available to treat them.

There are however studies that also apply machine learning to the admission of elective patients. In the study conducted by Nelson et al (2019), the authors use machine learning to assess whether or not patients will actually attend their scheduled appointments in hospital. Despite the resource being prepared for these patients, a no-show will result in a waste of this resource and this work aims to provide a way to then re-direct that resource. The authors use information on the history of the patient with a gradient boosting machine to get a strong predictive accuracy. Srinivas and Ravindran (2018) carry out this same prediction of no-shows to elective appointments, however they then leverage the risk of no-show in order to update the scheduling system of the hospital.

With many health systems providing long waiting times for appointments (Xavier 2003, Dimakou 2013), another important factor when it comes to elective patients is prioritising patients in the schedule. Yousefi et al (2019) approach this by first using a clustering algorithm to group patients into different priority categories. They then treat the schedule as a Markov decision process where waiting time for patients in the high priority clusters is to be minimised.

These approaches can be difficult to validate due to their direct impact on the scheduling of appointments. As a result, there is no chance to verify if the patients turn up or not once the schedule is changed. They also rely on historical behavioural data (such as how many times a patient has missed an appointment before) which are not stationary distributions and therefore limit how successful supervised learning can be in this domain in the long term.

Overall, the application of machine learning to predicting emergency patient admissions and scheduling is well-explored. Works are generally split between emergency and elective patients with further subdivisions according to the data used, the models used and what is being predicted (see figure 2). Very few works validate their models in hospitals in real-time, most using a retrospective test-set to assess performance. Furthermore, some models are difficult to validate due to being designed to intervene in the admission and scheduling process.

There is also very little connecting these studies. Most work is carried out with the data from the hospital that the authors are associated with and built around that. Due to hospitals being different, that leaves little scope for building on previous work or developing models that can be used universally. A public dataset that could be held as the gold-standard for patient flow would aid in this significantly as a benchmark for experiments.

Many of the studies that are created in predicting admission to hospital focus on subsets of patients with certain conditions. As these patients will require the same treatments and specialist staff to treat them, this can be seen as resource prediction for patients being admitted to the hospital from the ED.

An example is in Ong et al (2012) where heart-rate variability in the ED is used alongside other demographic information on the patient as input features to a support vector machine. This is then used to create a score on the likeliness of cardiac arrest occurring in the next 72 h. While this is not strictly framed for patient flow, this prediction allows clinicians to plan for resource in the cardiac department. Predicting whether or not a patient is septic is also important for patient flow in terms of resource planning. As a result, models predicting whether or not ED patients are suffering from sepsis have been developed (Horng et al 2017, McCoy et al 2017, Delahanty et al 2019). The authors use a mixture of information available at triage, demographic information and free-text to make prediction of whether or not the patient is septic, which if accurate, could allow planning of their treatment before the patient becomes critically ill.

In fact, there have been many such studies predicting whether or not a patient is suffering with a certain condition in the ED which allows resource planning. These include predicting if a patient is suffering from acute kidney injury (Martinez et al 2020), requires intensive care (Fernandes et al 2020, Finkelstein 2020), is suffering from a urinary tract infection (Taylor et al 2018), have bacterial infections (Ramgopal et al 2020) as well as predicting emergency hospitalisation of patients undergoing chemoradiation (Hong et al 2018).

While these predictions are useful for planning patient flow, they are not explicit predictions of admission. A more explicit approach is seen in Luo et al (2019) where the classifier is trained to predict admission to hospital of patients suffering from bronchiolitis.

While predicting admission to hospital from the ED is useful, a greater level of granularity, such as which departments in the hospital the patient will be admitted to, is more useful to clinical staff. An example is seen in Lee et al (2020) where rather than predicting admission, they predict the disposition of the admitted patient, choosing out of intensive care units, telemetry units, general practice units and observation units. As these 'ward types' tend to have separate resource, they are better able to adapt their resource according to the predictions made. This approach is also seen in El-Bouri et al (2020) where the authors also classify into 'ward types' to provide a similar level of granularity to the hospital admission prediction problem. However, in this case they use medical, cardiac, neuro, trauma, intensive care, surgical and general/obstetrics and gynaecology as their ward groupings. They develop a novel 'interpretable' layer for their deep neural network to guide information collection at triage and train the model using curriculum learning. El-Bouri et al (2020) further augment their model by using reinforcement learning to allow an agent to carry out the curriculum learning that maximises the performance of predicting where in the hospital patients will be admitted to. In order to make as general a model as possible, these studies of patient disposition do not consider subsets of patients but rather the entire population of the ED to replicate daily working conditions.

Figure 3 shows the general structure of works that have been conducted on predicting flow from the ED to hospital. It should be noted that as all of these works consider flow from the ED, all patients considered are emergency patients. Table 2 shows how readily available labelled datasets are for the EDii prediction problems and the popular approaches to tackling them. It should be noted that readily available here means data that would generally be saved on a hospital EHR and not data that would be easily accessible on a public dataset.

Zoom In Zoom Out Reset image size

Table 2. Popularity of different methods and data availability for each of these problems.

	EDii problem
	Hosp. admission	Hosp. admission loc.	Resource req'ment
Labelled datatset readily available?	$\checkmark$	$\checkmark$	$\checkmark$
Regression methods popular?	✗	✗	✗
Classification methods popular?	$\checkmark$	$\checkmark$	$\checkmark$
Bayesian methods popular?	$\checkmark$	✗	✗

We have seen in this section that the emergency-inpatient interface in hospitals, while being ill-defined in practice, is well researched using machine learning. Authors predict admission from the ED in order to provide information for clinical staff to prepare space should it be needed. To improve the performance of the classifiers, many authors condition their models on the demographics of the patients (e.g. elderly or young patients) or on the patient disease (patients suffering from the same ailment in the ED). In order to provide greater granularity on which resource will be used in the hospital, some authors also predict which 'ward type' will be used by the patient to be admitted to the hospital.

However, once again there is little connecting these studies. None of the studies reviewed build off each other or use the same dataset for comparison. Furthermore, the definitions used to categorise patients vary by paper. As was seen when categorising elderly patients some studies use 70 and over and some use 75 and over. Clearly it would be beneficial to have an agreed range to make models more comparable. This further emphasises the need for a shared, publicly available dataset for use when creating machine learning models for patient flow. All definitions of demographics should be included in the dataset so that researchers make valid comparisons to models. It will also be beneficial in allowing researchers to compare their methodologies and validate them on the same dataset as others as well as apply them to their own hospital's data. This will also make research more consistent, allowing researchers to build and improve upon each others models instead of applying similar models to similar problems using different data.

The most common way in which machine learning is used to provide predictions for 'inpatient flow' is through predicting if patients will be transferred to another ward. Note that while in section 6.1 we considered studies which investigated patient degradation as a signal for resource preparation, we will not consider degradation for inpatients as a signal for resource prediction. This is due to hospitalised patients generally being admitted to wards that are capable of handling patients in their condition. It is also due to the fact that using machine learning for the monitoring of inpatients for degradation has a very rich literature and would require a review of it is own (Clifton et al 2015). As a result, we only focus on works that explicitly predict admissions or transferrals of patients.

7.1.1. ICU transfer

During a patient's stay in hospital, various tests may be requested to help clinicians gain a better understanding of the patient's condition. These tests are also an important part of the patient flow process and timely testing helps to improve flow through the hospital. An example is seen in Molaei et al (2016) where the authors investigate whether or not they can predict if inpatients with traumatic brain injury require a CT scan using 'cost sensitive' random forests. In doing so, they aim to create a prioritisation system for scanning, which will allow faster treatment of patients and therefore a better patient throughput.

Another way in which resource management has been tackled with machine learning is in the scheduling of laboratory samples that need to be processed (Williams et al 2019). Again, by scheduling these samples in an efficient way, this allows patients to be treated more quickly in the hospital, and in some cases prevents the unnecessary hospitalisation of a patient.

These examples can be seen as assessing the risk of resource utilisation on a patient-by-patient basis. A more high-level view is used in Vieira and Hollmén (2016) where all resource is pooled together (anything including staff or use of machinery). Random forests are used to perform regression on the expected resource use in the next 30 d. While this has limited use to clinical staff due to the lack of granularity, it may be useful for budgeting purposes.

There are very few works that seek to predict the full patient journey through a hospital using machine learning. This may be due to the fact that transfers of inpatients is generally quite rare due to most inpatients being admitted to a ward that is capable of providing the appropriate care for them. Xu et al (2017) treat the hospital journey as a point process. They use a generalised linear model to predict the next location a patient will be transferred to as well as the dwell-time in that unit. They utilise the MIMIC-III dataset (Johnson et al 2016), which is an ICU based dataset and so the transitions they predict are between various types of intensive care unit. However, in terms of predicting the inpatient journey, this is a promising direction. Expanding to the entire hospital, it is possible to predict movement of patients between wards as well as for the use of machinery. Also predicting the dwell-time will allow for better planning of the flow of patients.

Of the four parts of the hospitalisation process that we have defined, the inpatient setting is the one in which machine learning has been used the least. The majority of studies investigate the transferral of inpatients to the ICU due to the resource-intense nature of ICUs. There have also been limited attempts at utilising machine learning to predict the expected resource that will be required by a hospital, either as a whole, or on a patient-by-patient basis. Very few works again have attempted to predict the whole hospital journey using machine learning. A common inconsistency throughout the literature is the prediction lookahead time that is considered. Standardising the lookahead time will allow studies to be more comparable and again, crucially, build upon previous work to further improve and integrate the field.

As the vast majority of studies are conducted with a clinical need in mind, this may reflect that the inpatient journey is not seen as a very important part of the patient flow problem. Figure 4 shows the structure of the studies that have been carried out in this area of patient flow. Table 3 shows the data availability for these prediction problems and popular methods used to tackle them.

Zoom In Zoom Out Reset image size

Table 3. Popularity of different methods and data availability for each of these problems.

	Intra-hospital prediction problem
	Ward (re)admission	Transfers	Resource forecast
Labelled datatset readily available?	$\checkmark$	$\checkmark$	$\checkmark$
Regression methods popular?	$\checkmark$	✗	$\checkmark$
Classification methods popular?	✗	$\checkmark$	$\checkmark$
Point processes popular?	✗	$\checkmark$	✗
Bayesian methods popular?	$\checkmark$	✗	✗

Discharge from ED has been treated as a classification as seen in Rahman et al (2020). The authors predict if a patient will be in the ED for longer than 4 h or not. They use features that are available early in the ED process to train a decision tree binary classifier. This approach is mimicked in Sariyer et al (2019) where various learning algorithms are experimented with to classify patients according to their length of stay in the ED. Azari et al (2015) acknowledge the large imbalance there tends to be in LOS datasets (with far fewer patients having long LOS), and present an ensemble method combined with multiple logistic regression to overcome this imbalance. However in this work they define a long stay as patients in the ED for longer than 14 h.

Rather than classify patients according to their likely LOS category, some authors prefer to use regression to predict each patient's LOS in the ED. Combes et al (2014) use linear regression model to predict the likely LOS of each patient presented to the ED. Ding et al (2010) instead use quantile regression but once again for the prediction of LOS in the ED. Feedforward neural networks have also been used for regressing the likely LOS of patients (Gül and Güneri 2015). One advantage to this approach of regressing the probable LOS is that there are no longer inconsistencies between studies on what is defined as a long-stay. However, this approach is also more difficult to train and achieve an accurate model in practice.

Predicting the LOS of patients in the inpatient setting is significantly more popular as a research area than in the emergency setting. This may be due to a prediction of LOS in the ED being less actionable than in the hospital where preparations can be made to ready a patient for discharge.

A hospital-wide approach is adopted in Pendharkar and Khurana (2014) where a regression tree is used to predict the LOS of patients admitted to hospitals in Pennsylvania using data that is available at the time of admission. This approach is also applied in Tanuja et al (2011), this time using a feedforward neural network to regress the LOS. These predictions are carried out at the time of admission. An alternative approach is to implement a classifier every day before discharge and predict the patients who can be prioritised for discharge as seen in Barnes et al (2016). In framing the problem in this way the authors exploit a static model for a dynamic problem by repeatedly applying the algorithm prior to discharge sessions at the hospital. They use a classification decision tree to prioritise patients ready for discharge.

Predicting discharge has also been approached as a time-series problem. In McCoy et al (2018) an autoregressive integrated moving average model is used to incorporate a time-series of seasonal data to predict hospital discharge volume. They compare this with using Prophet (Taylor and Letham 2018), an additive regression model developed by Facebook Research for forecasting seasonal trends, for the same task. An NLP approach has also been used where the clinical notes from the ED are used in order to predict if a patient will be admitted to the hospital for more than 2 d (Bacchi et al 2020).

As has been a common theme throughout this review, discharge predictions are also conditioned on patient demographics. In other sections this is primarily to improve predictive performance amongst patient subgroups. However, in discharge prediction this is due to certain patient subgroups being more likely to be 'bed-blockers' such as elderly patients (Launay et al 2018). To maximise clinical utility it is more effective to condition the training dataset on these subgroups and apply the algorithms to these patients only. An example is in Elbattah and Molloy (2016) where a regression forest is used to predict the LOS of elderly patients in a hospital and a random forest is used to predict the location of discharge for these elderly patients. These predictions are used in conjunction with a discrete-event simulation in order to simulate the flow through an Irish hospital. Children are also a cohort of patients in which there can be great variability in LOS. To address this, Castiñeira et al (2020) use a gradient boosted tree to classify whether or not a child will be a long-stay patient in the paediatric ICU (with long-stay being defined as a stay of greater than 4 d). They also use the static model for a dynamic problem approach by extracting features from the time-series of the patient's vital signs and repeatedly feeding these to the classifier. Note that this prediction concerns the LOS within a ward and not the hospital stay as a whole.

As with conditioning on demographics, conditioning on co-morbidities is also done in discharge prediction. In fact, this tends to be the most popular form of setting the problem due to patients with different ailments and treatments generally requiring different recovery times.

One such prediction is carried out for patients with congestive heart failure in which the authors apply a static cubist model (Quinlan 1998) dynamically as data is updated during the patient stay (Turgeman et al 2017). The model is used to regress the likely LOS in hospital of the patient.

Further discharge predictions have been carried out on patient cohorts who have suffered from stroke (Al Taleb et al 2017), patients who have suffered hip-fracture (Elbattah and Molloy 2016), patients suffering from schizophrenia (Kirchebner et al 2020), patients admitted for cardiac care (Daghistani et al 2019), patients post-brain tumour surgery (Muhlestein et al 2019), patients who have undergone total hip-arthroplasty (Ramkumar et al 2019) and patients who have undergone surgery due to colorectal cancer (Stoean et al 2015). In all of these studies, there is no consensus for defining a 'long-stay' patient.

Discharge prediction is one of the more popular areas of patient flow for researchers to apply machine learning. Discharge prediction has been carried out by either predicting whether a patient is likely to be long-stay or by directly regressing the expected LOS of the patient. It has been applied to both emergency and inpatient settings. In the inpatient setting, studies have conditioned their datasets according to demographic. There have also been studies that condition their dataset according to the comorbidity or treatment that the patients of interest have undergone.

A clear inconsistency between studies is the definition of a long-stay patient. Having a common dataset with pre-defined long-stay patients will improve the ability of researchers to compare models and build upon previous work. Figure 5 shows the structure of the literature published in this field. Table 4 shows data availability and popular methods used to tackle the discharge problems. It should be noted that the difficulty with a labelled dataset for discharge readiness is that generally it is not recorded when a patient is ready for discharge but when they actually are discharged.

Zoom In Zoom Out Reset image size

Table 4. Popularity of different methods and data availability for each of these problems.

	Discharge prediction problem
	Hospital LOS	Long-stay prediction	Discharge ready
Labelled datatset readily available?	$\checkmark$	$\checkmark$	✗
Regression methods popular?	$\checkmark$	✗	✗
Classification methods popular?	$\checkmark$	$\checkmark$	$\checkmark$
Point processes popular?	✗	✗	$\checkmark$
Bayesian methods popular?	✗	✗	✗

While all of the problems outlined in the above review are important for clinical practice, solving some of these problems is more urgent than solving others. An example of a high-priority problem to solve is predicting readiness for discharge. One of the greatest problems dealt with in patient flow is the 'bed-blocker' phenomenon whereby patients do not have appropriate destinations to be discharged to. Predictions of readiness for discharge will not solve the lack of space in care homes, however it will allow for more effective allocation of the time and attention of clinical staff.

An equally important task is the prediction of ED admissions. This represents the front-end of the hospital with the discharge readiness representing the back-end. Being able to accurately predict patient admissions numbers in the ED would allow for accurate planning of staffing rotas thereby reducing costs and time wasted. It would also greatly improve the care provided for each individual patient.

Following on from this, should these predictions not be accurate enough, solving the ED-inpatient interface problem would be the next most important. This prediction would prevent the filling up of the ED due to inability to transfer patients into the hospital. Having an accurate model here would create a more streamlined flow of patients into the hospital, but naturally would depend on there being enough flow out as well.

Finally, the problem that should be least prioritised is inpatient transfer prediction. Despite being important, inpatient transfers are generally quite rare due to patients being admitted to appropriate wards from the outset. However, there is value in predicting resource flow and patient movement in order to plan that resource.

9.2.1. Data challenges

9.2.2. Technical challenges

The majority of studies discussed in this review take advantage of the fact that there exist EHR systems in many modern hospitals which allow data extraction and dataset creation. However, there remain challenges in terms of data collection for the different tasks at hand.

The ED admission prediction relies on seasonal information which can be correlated with admissions but is generally a difficult prediction to make. Wearable sensors could benefit this prediction greatly, providing more granular information to the hospital. The sensors could also be provided to patients who need them most (and are most likely to be brought to the ED in an emergency such as elderly patients in care homes).

We believe that further improvements to data collection could be made in the inpatient journey as well as in discharge prediction. Currently, while scans in the hospital are logged on the EHR, the movement of patients to scans are not and nor is the resource associated in moving that patient. These data would be very helpful to provide a more complete picture of what resource each hospitalised patient utilises and thereby helping machine learning scientists create more accurate predictions of the likely resource needed.

In discharge prediction, one of the challenges is that it is generally not recorded when a patient is medically ready to leave the hospital but when they actually do. Augmenting a dataset with this information could help predict when a patient is ready to leave hospital and in doing so, allow the team looking after them to move their resource to more vital care, with a more generalised team looking after the patient thereafter until discharge.

The first research direction to be considered is multitask learning, a machine learning method that allows multiple tasks to be learned at the same time. One of the aims is to exploit the learning signals generated by training on one task to create an inductive bias in the model that will allow the effective learning of another task by the same model (Caruana 1997).

Multitask learning can be applied to different problems across the four domains of patient flow to both related (e.g. predicting risks of various in-hospital complications) or unrelated tasks (e.g. predicting length of stay in the ED and predicting hospital admission destination). Once again this relates to the usefulness of having more granular information for clinicians to work with. An example may be when predicting the location of admission of a patient to hospital, also having some prediction of whether the patient is likely to deteriorate or not. This gives better indications of the likely resource requirement of the patient as well as their likely trajectory within the hospital. While this could be done using separate models for each prediction, a single model that can embed an accurate representation of the patient will be more informative and useful to clinical staff. As a result, a key component of this work will be in the development of representation learning algorithms (Bengio et al 2013, Van Den Oord et al 2017) that are capable of representing patient conditions upon presentation to the ED or admission to hospital.

Multitask learning has been applied in many healthcare applications to leverage the shared information across different tasks. Huang and Dong (2018) have used multitask learning to predict major adverse cardiac events, identifying each type of adverse event as a single task as opposed to having a multiclass classification. Xia et al (2019) have also used this approach to predict prescription patterns for various drugs that are given to similar patients. Multi-task learning has also been used in medical imaging. Khosravan et al (2019) have used multitask learning in the detection of abnormal nodules on chest CT scans for lung cancer screening. They jointly train their model to segmenting potential abnormalities and identify the presence of a nodule in the region of interest. This is further evidence of how more granular information from the model can provide clinicians with better insights into the condition of the patient.

Transfer learning is based on the principle of knowledge transfer across different machine learning tasks and models. It is based on the notion that knowledge gained by the algorithm when trained to solve a particular problem can be stored and applied to solve another related problem, which means it is closely related to multitask learning. This approach includes transferring knowledge from the source domain, D_S , to the target domain, D_T , to help improve the learning of the target-domain task, T_T .

Transfer learning can provide significant advantages in the applications of machine learning in patient flow. It can enable (a) the transfer of knowledge across different tasks and (b) the transfer of knowledge across different populations. The former can help overcome the lack of clinical data for certain problems. For example, one of the barriers to developing effective machine learning tools for COVID-19 patients is the lack of data on COVID-19 patients. A transfer learning approach can provide a solution by using a model that is pretrained on a large non-COVID-19 dataset and adapting it to perform the task of interest in COVID-19 patients. Transfer learning has been used to overcome the lack of COVID-19 imaging data by Mahmud et al (2020). They trained a convolutional neural network by using a pretrained a neural network (pretrained on a dataset of bacterial and viral pneumonia chest x-ray scans) and fine-tuned this using scans from COVID-19 patients. This was done due to the scarce availability of chest x-rays from these patients.

Transfer learning can also help us transfer knowledge across different populations. This is valuable clinically given the diversity and differences in the genetic predispositions, prevalence of diseases, lifestyles, and risk factors across different populations. Mao et al (2018) used transfer learning to generalise their sepsis prediction algorithm to a new healthcare setting. They trained their prediction model using data from the MIMIC-III dataset (data from ICU patients) and transferred the model to a dataset from the University of California, San Francisco (UCSF) Medical Centre (a dataset of in-hospital patients from a variety of specialty wards). Their transfer learning approach was based on adding incremental amounts of data from the UCSF dataset to the MIMIC training dataset, resulting in better generalisation of the model to the dataset being introduced.

Transfer learning represents an interesting target for future research in patient flow machine learning applications. Transfer learning can be used to generalise models across different healthcare contexts and to overcome a lack of recorded data.

Continual or lifelong learning refers to the ability to continually learn over time by accommodating new knowledge while retaining previously learned experiences. This approach has the potential to enable machine learning models in the healthcare space to adapt and adjust automatically to new context and settings like a new healthcare context, new patient population, or a new and emerging disease. This has the potential to enable the creation of dynamic clinical AI models that optimise clinical management decision in real time and learn from the continuous influx of information in real world healthcare context. A continually learning algorithm should be an adaptive algorithm capable of learning from a continuous stream of information, with such information becoming progressively available over time. The accommodation of new information should occur without catastrophic forgetting or interference (Parisi et al 2019).

However, continual learning represents a long-standing challenge due to the susceptibility of machine learning models to catastrophic forgetting. This phenomenon refers to the decrease in model performance or the complete overwriting of the previously learned information when new knowledge is introduced.

A paper published in the Lancet in 2020 (Lee and Lee 2020b) highlights the promise of continual learning in revolutionising the applications of clinical AI and leveraging the continuous influx of clinical information to improve patient care. Shah et al (2019) highlight that machine learning algorithms that are capable of continuous learning are a critical future research and translational direction in healthcare AI. They also report that the FDA is considering widening its regulatory framework to include AI-based Software as Medical Device (SaMD) systems that are capable of continuously learning and optimising performance in real-time to improve patient care.

Continual learning promises considerable value in patient flow as it would enable machine learning models to adjust to different healthcare settings continuously and automatically. Therefore machine learning algorithms would be able to absorb the variation across different healthcare institutions and patient populations. Moreover, continual learning may enable machine learning algorithms to continuously learn after deployment to clinical settings gradually improving their performance through use.