Development and implementation experience of a learning healthcare system for facility based newborn care in low resource settings: The Neotree

Abstract Introduction Improving peri‐ and postnatal facility‐based care in low‐resource settings (LRS) could save over 6000 babies' lives per day. Most of the annual 2.4 million neonatal deaths and 2 million stillbirths occur in healthcare facilities in LRS and are preventable through the implementation of cost‐effective, simple, evidence‐based interventions. However, their implementation is challenging in healthcare systems where one in four babies admitted to neonatal units die. In high‐resource settings healthcare systems strengthening is increasingly delivered via learning healthcare systems to optimise care quality, but this approach is rare in LRS. Methods Since 2014 we have worked in Bangladesh, Malawi, Zimbabwe, and the UK to co‐develop and pilot the Neotree system: an android application with accompanying data visualisation, linkage, and export. Its low‐cost hardware and state‐of‐the‐art software are used to support healthcare professionals to improve postnatal care at the bedside and to provide insights into population health trends. Here we summarise the formative conceptualisation, development, and preliminary implementation experience of the Neotree. Results Data thus far from ~18 000 babies, 400 healthcare professionals in four hospitals (two in Zimbabwe, two in Malawi) show high acceptability, feasibility, usability, and improvements in healthcare professionals' ability to deliver newborn care. The data also highlight gaps in knowledge in newborn care and quality improvement. Implementation has been resilient and informative during external crises, for example, coronavirus disease 2019 (COVID‐19) pandemic. We have demonstrated evidence of improvements in clinical care and use of data for Quality Improvement (QI) projects. Conclusion Human‐centred digital development of a QI system for newborn care has demonstrated the potential of a sustainable learning healthcare system to improve newborn care and outcomes in LRS. Pilot implementation evaluation is ongoing in three of the four aforementioned hospitals (two in Zimbabwe and one in Malawi) and a larger scale clinical cost effectiveness trial is planned.


.]
Conclusion: Human-centred digital development of a QI system for newborn care has demonstrated the potential of a sustainable learning healthcare system to improve newborn care and outcomes in LRS. Pilot implementation evaluation is ongoing in three of the four aforementioned hospitals (two in Zimbabwe and one in Malawi) and a larger scale clinical cost effectiveness trial is planned.
behavioural sciences, global health, health services, neonatal 1 | BACKGROUND AND RATIONALE Worldwide, 2.4 million children younger than 28 days die every year representing 48% of deaths in children under 5 years 1 and at least a further 2 million are stillborn. 2 About 90% of newborn deaths and 98% stillbirths occur in low-resource settings. 2,3 Most mothers (~80%) now deliver in health care facilities. 4 However, this increase in facilitybased deliveries is not associated with expected reductions in maternal and newborn mortality. 4 Up to 70% of newborn deaths and at least half of stillbirths are avoidable through the consistent implementation of low-cost evidence-based interventions. 5,6 Most recent WHO newborn data indicate coverage, quality, and measurement gaps in newborn care. 7 Health systems strengthening, along with education and training in newborn care are key to saving newborn lives. [7][8][9] Previous studies have demonstrated the value of educational interventions for healthcare professionals and kangaroo mother care in decreasing newborn case fatality rates. The potential for e-health technologies and telemedicine to improve newborn care is increasingly being demonstrated. 10 These digital formats can provide a user-friendly interface for the implementation of evidence-based interventions and guidelines, reliable data systems, digital clinical decision support tools and education in one platform. 11,12 However, even when strong evidence of effectiveness exists, intervention coverage is often low due to lack of IT-skills training, human resources and finance, 4,13 as well as lack of co-development, government buy-in and alignment with existing country systems. 14

| Reliable capture of routine healthcare data in low resource settings
The capture of routine health data through Electronic Healthcare Records is key to improving quality of care, efficiency and cost-effectiveness, and a crucial building block for strong health systems. 7 Yet, few such Electronic Healthcare Record systems exist in low-resource settings 15 and where they do, they have historically been focused on specific disease processes such as HIV. Furthermore, data capture and storage systems are predominantly retrospective and paper-based, making it inefficient to retrieve and use the data to inform care decisions. Weak health information systems, especially gaps in the provision of reliable, accurate disaggregated and timely data to guide decision making have been highlighted as a significant barrier to equi- These decision support tools provide healthcare professionals with targeted information for a given patient or situation. Healthcare professionals enter data at the point-of-care and those data generate case-specific advice according to evidence-based guidelines (commonly known as knowledge-based systems 16 ). Digital knowledge-based clinical decision support tools have been shown to improve diagnosis and treatment decisions. For example, implementation of a sepsis clinical decision support tool in a US hospital was associated with a 53% reduction in adult sepsis-related mortality. 17 Few low resource healthcare systems have adopted knowledgebased clinical decision support systems at this stage. This is despite most small and sick babies (an estimated 30 million worldwide) seeking facility-based care within these settings. 18 Knowledge-based clinical decision support applications have been trialed in Tanzania and Malawi for the community case management of maternal and childhood health conditions. 19,20 A digital clinical decision support addressing a narrow range of newborn conditions (Noviguide) linked to education in newborn care (but not to Electronic Health Records) has been piloted in Uganda. 21 A prototype decision support app is under development in Kenya with pilot data pending. 22 Digital support tools have huge potential to improve clinical outcomes in these settings. However, they rely on robust data capture systems, and evidence for impact and scale up is lacking. Furthermore, there are many gaps in the evidence with which to create clinical guidelines that are relevant to low-resource settings. 23 More recently, non-knowledge based clinical decision support tools are being developed based on adaptive prediction models, including machine learning methods. Examples of how this approach has been successfully applied in child health in high-resource settings include Longsdale 2020. 24 Machine learning has shown improvements in predictive models for assessing the need for critical care and risk of mortality on admission to paediatric intensive care units. 25 Limiting factors in the use of machine learning to optimise data-driven predictive clinical models are poor quality of data, limited understanding of the platforms they are delivered into, and deficiencies in their implementation into practice.
1.3 | Education in newborn care to improve quality of care in low-resource settings Education in newborn care is essential to upskill healthcare professionals in delivering evidence-based practice. 9 Healthcare providers have cited lack of nursing and medical training in the provision of higher-level neonatal care as a barrier to providing quality in-hospital care to newborns. 26,27 Additional barriers to education in newborn care include high staff turnover and frequent reassignment of staff to different units within the facility. Basic newborn care educational programs have been shown to improve knowledge, competence, and appropriate practices regarding newborns. However, not all basic education packages result in a sustained change in practice 22 and reliance on paperbased reporting systems has hampered implementation. Neotree has been developed as a solution through its data capture and constant reinforcement of education messages at the point of care. 28 Digital approaches to education in newborn care are being explored. 21,22 However, these have not been combined with, and linked to, electronic health records and clinical decision support systems; and sustained improvements in quality of care have yet to be demonstrated.

| Learning healthcare systems
In high income countries, data capture, decision support, education, and continuous learning have been combined to create learning healthcare systems to accelerate health system strengthening and performance. The concept of a learning organisation 29 was first applied to healthcare systems in the United States (US) in 2007 as a way of leveraging electronic health record data to rapidly develop evidence for daily clinical practice and policy. 30 In learning healthcare systems, data are collected and collated from multiple sources, for example, Electronic Health Records and patient experience. These data are then analysed and interpreted to create knowledge and evidence, for example, optimising existing evidence based clinical guidelines. This knowledge is then fed back into the healthcare system to improve health care and outcomes through a combination of automated delivery of knowledge, such as, through digital clinical decision support tools and education, quality improvement and implementation science.
Quality improvement aims to systematically improve and monitor the quality of care, for example, feedback via electronic data dashboards. Implementation science aims to understand and reduce gaps between what is known (evidence) and how knowledge is translated into practice (behaviours) through various strategies, such as Audit and Feedback. The Audit and Feedback strategy motivates health professionals to improve their practice by visualizing and highlighting the gap between their own performance and desired performance targets.
Ideally, learning healthcare systems promote an iterative, synergistic cycle of interaction between data, knowledge, and practice delivered on an integrated platform, compatible with local systems and culture resulting in a constant state of quality improvement. Learning healthcare systems also offer the capacity for real-time learning including quasi-experimental designs, to greatly increase the ability to generate and test hypotheses in a timely manner. A learning healthcare system can exist at any scale be it facility, national or international.
Patients, family, and community engagement and the assurance of high standards of data quality, governance, and accessibility are all central to the delivery of a successful learning healthcare system.
A scoping review of global learning healthcare systems described 68 such systems; the majority of these were in the US, two in the UK and only one in a low-resource setting, Kenya. 31 One of the seven recommendations of a recent commission into the future of the UK NHS was to develop the culture, capability, and capacity to become a learning health system. 32 In the US, the "ImproveCareNow" network is a mature learning health system for child health aimed at improving health outcomes for children and young people with inflammatory bowel disease. 33 This network has demonstrated improvements in remission rates and growth through standardized data collection, monitoring, evaluation, as well as sustainable and collaborative care.
Despite low-resource settings have the most to gain from learning health systems, such an approach is uncommon, perhaps due to a lack of knowledge, research, or logistical capacity. 34 The 2016 WHO framework for improving health facility newborn care highlights the need for actionable information systems. 35 Delaying the development of learning healthcare systems in low-resource settings until they can be embedded within fully functioning healthcare systems will only exacerbate existing health inequities. and Zimbabwe (2018 onwards) to co-develop a learning health system for newborn care in low-resource settings -the Neotree. We have also conceptualised a similar, linked perinatal learning healthcare system -Mummytree.
Throughout the process, we have used open-source code and maintained local data ownership. 28,36 The Neotree system combines an android application with accompanying data visualisation, linkage, and export. Its low-cost hardware and state-of-the-art software support healthcare professionals to improve postnatal care at the bedside. Neotree is a horizontal intervention -aiming to comprehensively address most common newborn disorders as opposed to focusing on one disease. It has been, and can be, readily adapted to incorporate new disease trends in outbreak situations such coronavirus disease 2019 as (COVID-19). 37

| Logistics of use
Healthcare professionals capture clinical and demographic data on newborns on admission, discharge and from laboratory data using low-cost android tablets at the bedside (see Figure 1 data pipeline; Figure 2 sample screen shots of the data capture screens). In our country settings (Malawi and Zimbabwe) neonates are usually admitted and discharged by nursing cadres. Hence Neotree has been developed for nurses by nurses. However, in one of our sites (Sally Mugabe Central Hospital) junior doctors take responsibility for admissions and discharges; hence in this site Neotree has been adapted for use by doctors.

| Data export and linkage
Neotree has been designed to work in low-resource settings where network connectivity can be limited Tablets   An open-source development process means the software can be easily adopted by hospitals/governments who can then tailor data input fields locally without needing remote software support or creating commercial dependencies. 36 In terms of open data curation and management, for example, we work in partnership with the Zimbabwean Ministry of Health and Child Care Electronic Health Records team to ensure seamless integration of Neotree data into the F I G U R E 1 Data pipeline: Currently, the web-based editor platform is used to design five NeoTree data and decision support forms. HCPs capture data, print to patient notes and export pseudonymized data to a data server. Data are processed, merged and analysed before being presented back to HCPs in meaningful dashboards. Data are also integrated with national electronic medical records F I G U R E 2 Sample screen shots of NeoTree front end (App interface): Recorded observation, examination, and data capture by the HCP triggers timely guidelines, education, and management giving decision support. On completion, captured data are automatically exported ready for analysis and presentation national system, with all data to be held and owned on a Zimbabwean Ministry server.
We are committed to anonymised open data sets being available for in-country policymakers and clinicians/academics to improve neonatal health outcomes. We follow the FAIR data principles to optimise use and re-use of our code. 39 The Neotree and Mummytree have been, and are being, developed to be compatible with local information technology systems and data regulations. This compatibility has been country specific according to country data regulations and preferences. In Zimbabwe, all data are stored on a physical server resident in Zimbabwe. In Malawi, data are stored in an encrypted storage area of the Amazon Web Server. In Zimbabwe, the data export and linkage to the Ministry of Health EHR has been configured precisely to match the national EHR.  Currently the dashboard includes two reporting interfaces for the Neotree system. The first dashboard displays real-time reports on a screen display in neonatal wards visualising key data (eg, admission and mortality rates). The second dashboard is a slide-deck of monthly data available for presentation at local morbidity and mortality meetings. 41 A third targeted QI screen for improving key clinical indicators (eg, admission hypothermia) has been prototyped and piloted pending implementation in the future. 38 The automated delivery of morbidity mortality data to local multi-disciplinary teams is crucial for clinical F I G U R E 3 Data dashboard development, Kamuzu Central Hospital, Malawi: Moves units from hand-drawn, time consuming, unreliable and unclear data charts to clear, dynamic digital displays of meaningful data insights and driving impactful interventions (eg, hypothermia highlighted in lower right dashboard) 38 decision-making, resource management, and the monitoring of quality care and standards. However, recent data from the WHO quality of care network show only 17 of 47 countries in the Africa Region undertake these audits. 7 The final route employed to drive clinical practice is through quality improvement methodology, used by health providers to iteratively improve and monitor care. 42 Figure 4 shows the draft logic model. In 2018 Neotree-beta was implemented in Sally Mugabe Central Hospital in Zimbabwe (annual delivery rate~12 000) as a Quality Improvement tool for neonatal sepsis. 42 Discharge data capture and laboratory data capture functions were added. (annual delivery rate~4500) to create Neotree-beta MVP-2. 38,41,48 A focus of this work was the evaluated co-development of a dashboard prototype as an electronic data feedback mechanism. 48 Acceptability of Neotree-beta MVP2 was found to be high with some feasibility issues raised in an evaluation using behavioural science frameworks. 45 Throughout this period the data linkage and pipeline features were iteratively developed.

| Clinical decision support
The clinical decision support function of the Neotree was developed in three parts. First, national and international standardised neonatal resuscitation and stabilisation guidelines were digitalised and incorporated into Neotree as a resuscitation algorithm, for example, if a baby was not breathing, advice was immediately given on how to resuscitate. 28 Second, a larger set of clinical decision support algorithms were developed according to Malawian national (Care of the Infant Newborn: COIN 50 ) and international guidelines and evidence in newborn care (Table 1). These were configured within the Neotree editor form of the application but not activated as we identified significant gaps in evidence-based guidance suitable for low-resource settings. For example, the European definition of neonatal sepsis is two or more clinical symptoms and two or more laboratory signs in the presence of, or because of, suspected or proven infection. 51 This definition is not possible in low-resource settings where laboratory investigations are not routinely available.
In the absence of extensive trial or epidemiological data in lowresource settings, alternative techniques to consolidate best available low-quality evidence can be used, such as expert opinion using the Delphi method.
In 2017 to 2018, we conducted a Delphi study to determine whether a panel of 22 neonatal experts with global expertise could address evidence gaps in four neonatal guidelines designed to be included in the Neotree: sepsis, neonatal encephalopathy, respiratory distress, and thermoregulation. 23 These conditions represent the leading preventable causes of neonatal mortality and are difficult to diagnose and manage appropriately in low-resource settings with some of the weakest WHO GRADE recommendations and quality of evidence.
Key changes made in response to this Delphi study were as follows.
First, the Thompson score, a validated sensitive clinical scoring system for diagnosis of neonatal encephalopathy in low-resource settings, 52 was adopted. Second, analysis was initiated to identify a set of triggers to prompt the healthcare professionals to carry out a Thompson score assessment, for example, resuscitation longer than 10 minutes after birth. 53 Third, additional work was commenced to devise and refine a sepsis risk score for low-resource settings (completion and integration into the Neotree anticipated summer 2022). In 2019, a scoping review of existing literature on clinical prediction models to diagnose neonatal sepsis in low-resource settings was performed. 54 In 2020, a dataset was constructed from the routine admission and discharge Neotree data from the neonatal unit of Sally Mugabe Central Hospital, Zimbabwe. A clinical prediction model to diagnose neonatal sepsis was then developed on this dataset by fitting multivariable logistic regression models. 55 The resulting prediction model is currently being refined on a second training dataset from Zimbabwe.
Fourth, all respiratory conditions are being placed under the umbrella diagnosis of respiratory distress of the newborn as experts concluded it was not easy in these settings without access to routine investigations to differentiate between causes of respiratory distress, for example, meconium aspiration vs congenital pneumonia. The remaining non-resuscitation algorithms are being refined according to best available evidence. All these IF-THEN knowledge-based decision algorithms will then be configured via the Neotree editor platform ready for testing. We are currently undertaking usability testing for the automated surfacing of the clinical problem list and linked management pages in response to entered data (rather than by healthcare professional choice). Once non-resuscitation algorithms have been finalised, we will undertake one-to-one usability workshops with exemplar clinical cases followed by implementation evaluation of acceptability and feasibility within the clinical workflow (February to April 2022).

| Neotree driving change in clinical care
To date we have gathered data for more than 18 000 babies, and over 400 healthcare practitioners have interacted with the Neotree system.
We have observed how the Neotree system can directly and rapidly change clinical care and strengthen adherence to evidence based clinical practice.
An example of such change in clinical care is to manage hypothermia, which is a preventable risk factor for poor outcomes: when the T A B L E 1 Clinical decision support algorithms generated by NeoTree (other than resuscitation guidelines) according to degree of complexity (simple vs complex) and level of underlying evidence (strong vs weak) Significant challenges were met during the initial pilot implementation study in Sally Mugabe Central Hospital, Zimbabwe, with capturing and linking microbiology laboratory data. 42 The need to ensure high quality Wi-Fi for intermittent data sync was identified as a second priority for ongoing implementation.

Key lessons learnt from the initial pilot implementation in Kamuzu
Central Hospital (May 2019-September 2019) work were several fold.
First, digital health interventions can be optimised combining both agile user-focused methodologies with behaviour change frameworks.
Second, that data on dashboards should be accompanied by clear messaging around how to act on those data, to motivate change in clinical practice (via audit and feedback). Third, complex data transformations should not be handled within the visualisation programme, but elsewhere in the system (eg, a data pipeline) to reduce network requirements in low resource settings.

| Summary and next steps
In this paper we have described our experience of developing and implementing a learning healthcare system for newborn care: Neotree. We are gathering ongoing implementation evaluation data and future steps include collecting robust evidence of clinical and cost effectiveness of impact on newborn care and mortality.
Ensuring adequate data quality and robustness, harmonization and integration with existing infrastructures are some of the key challenges to implementing novel healthcare systems. Therefore, we continue to review our collected data and develop measures to control data quality in the long-term while ensuring the application is user-friendly for healthcare professionals without being a burden to productivity.
Over the last year we have also been working with partners to strengthen our parent, family and community engagement with promising results. Work to progress the Mummytree is ongoing, and we are also working with Zimbabwean paediatricians to develop and test a strategy to link Neotree data with neurodevelopmental follow-up.
The very essence of a learning healthcare system is the ability to continuously learn from data and experience. This adaptive system requires a skilled workforce to support ongoing development and implementation. Therefore, future next steps will also include explicit building of capacity and capability in the clinical and academic workforce in low-resource settings to enable the sustainable development and delivery of Neotree and other similar systems.

CONFLICT OF INTEREST
Michelle Heys and Felicity Fitzgerald are both trustees of the Neotree charity (www.neotree.org) but receive no financial payment from this role. Caroline Crehan was a trustee of the Neotree charity (stepped down in 2018) and received no financial payment for this role. There are no other conflicts of interest to declare from any other co-author.