Using Machine Learning to Model Early-onset Neonatal Sepsis Risk in Uganda and Zimbabwe

There will be three main phases to this study, which is an observational cohort study.

Phase 1 will take place at Kawempe Hospital in Kampala, Uganda. Participants will be recruited from labour wards by study staff, and interim analysis will be completed after recruitment of 10% of the total sample. See section 7.3 for further details.

Data will be collected via an online CRF, including the factors listed in the table in Section 3.

Additionally:

• All neonates within the study will have a neonatal venous blood sample taken for culture by designated study staff
• This must be taken before administration of antibiotics or within 48 hours of birth, whichever is sooner, and regardless of whether the neonate is being investigated and treated for sepsis.
• Postnatal information will be collected via telephone at 1 month including maternal and infant outcome, hospital readmission or clinical illness, and vaccines given All data will be checked for quality control by a member of the study team. A blood culture will be taken from all neonatal participants at birth, prior to starting antibiotic treatment wherever safe to do so. The volume of blood taken for culture must be 1-2ml and will be verified by weighing blood culture bottles. Training will be provided to paediatric and study teams on blood culture sampling methods, including adequate cleaning of the skin prior to sampling. Feedback on sample volumes and contamination rates (see section 7.3 for definition) will be fed back to paediatric consultants, and additional staff training offered. Where sample is remaining after 3ml is added to the culture bottle, the excess will be retained as a blood spot sample for use in future ethically approved research, where the participant has consented to this.

The results of the microbiological tests will be communicated to the participant's clinical team in real time - see section 7.3.4 for further details. Clinician diagnosis of sepsis will be recorded where a clinician of registrar level or above has recorded a diagnosis of sepsis within 48 hours of birth. To reduce the subjectivity of this measure, GAIA level of certainty for a diagnosis of sepsis will also be assessed (35) and the level of diagnostic certainty recorded (1-5, as per GAIA criteria). Medical records are recorded on paper at Kawempe Hospital, before being coded into an electronic medical record by data clerks. PDF scans of medical records will be available where detailed review is required.

All participant data will be collected by trained study staff, and recorded into an online CRF hosted on REDCap, a secure web application. No patient identifiable data will be stored on the main REDCap database. See section 8.6 for further details of data management plans.

The postnatal information will be completed via a telephone call at 28 days post-delivery, made by study staff. Participants' contact number, a contact number for their partner, and a contact number for a second friend or family member will be taken at recruitment to minimise loss to follow up due to changes to mobile telephone number. Participant identity will be confirmed using three factors (date of birth, full name, address/village). The following information will be collected:

• Health status of mother and baby
• Any visits to medical care for the neonate
• If attended or admitted to Kawempe hospitals, these electronic medical records will be reviewed by study staff
• If attended elsewhere, the mother will be asked the reason for presenting, whether the neonate was admitted to hospital, the main diagnosis, and any treatments given.

Recruitment of the first 160 participants from labour ward (10% of the overall sample) will be treated as a pilot, and both recruitment and data will be reviewed at this point with particular attention to the following aspects:

• Demographic comparison between recruited women and the population attending Kawempe
• Data completion
• Volume of blood culture samples taken
• Loss to follow up rates
• Association between potential risk factors and clinical sepsis, or culture-positive sepsis After this pilot analysis, the data collection instrument may be adjusted to be more concise if some risk factors appear to have no relationship with neonatal outcomes. Additionally, the recruitment strategy may be adjusted to include women attending antenatal clinics to improve the representativeness of the sample if the pilot sample is significantly different demographically to the general population of pregnant women attending the same study site antenatally. Finally, feedback and staff training will be given as needed to ensure optimal sample volumes and data completeness.

During Phase 1, two separate focus group discussions with 10 participants each will be held with staff from Kawempe Hospital. One discussion will be with senior clinicians, and one with junior clinicians and midwives. The focus groups will be held on site at Kawempe in English. Focus groups are being held in order to complement the quantitative results of this project by highlighting risk factors currently perceived as significant (which should then be a key part of analysis and results dissemination to confirm their usefulness), inform of current diagnostic difficulties to guide future work with the risk stratification model, and provide context for development and communication of the model to maximise its utility. The focus groups will address the following issues:

• Ascertain knowledge and awareness of neonatal sepsis
• Management of neonatal sepsis
• regarding risks and consequences of neonatal sepsis, and
• Current diagnostic strategies and their perceived strengths and difficulties
• Perceptions of neonatal sepsis and how this might affect diagnosis and treatment
• Recommendations to enhance the diagnosis and treatment of neonatal sepsis The focus groups will be conducted by members of MUJHU's social sciences team and Dr Sarah Sturrock. A tool has been developed by the social sciences team and Dr Sturrock for these discussions (see attached). The discussions will be recorded, and recordings destroyed once the discussions have been transcribed. The social sciences team will then complete thematic analysis of the discussions.

The investigators anticipate Phase 1 will take place between January and July 2024.

Phase 2 will use the data collected in phase 1 to construct a risk stratification model, in collaboration with the Advanced Research Computing Centre at University College London.

The investigators anticipate Phase 2 will take place between March and September 2024.

Phase 3 will involve external validation of the risk stratification tool. This phase will take place at Sally Mugabe Central Hospital in Harare, Zimbabwe. Participants will be recruited from antenatal clinics and labour wards by study staff.

Similar data will be collected from participants in Phase 3 as in Phase 1; following Phase 2, the data fields required for the model will be determined and these fields will be collected for Phase 3 participants.

Data collection will take place via the Neotree platform, which is in use at Sally Mugabe(41). Phase 1 will collect data on items not part of a standard medical history (such as sanitation availability and food security). Should any of these items form a key part of the statistical model, the possibility of adding these items to the Neotree platform will be discussed with the Neotree team. Pseudonymised data only will be extracted from the Neotree dataset, and the model from Phase 2 applied to determine the risk of neonatal sepsis.

The primary outcome for Phase 3 will be senior clinician diagnosis of sepsis recorded within 48 hours of delivery, and blood cultures (when available) will only be taken in neonates judged to be at risk of, or with clinical signs of sepsis as per standard clinical guidelines. Using senior clinician diagnosis of sepsis as the outcome is because supply of blood culture bottles and laboratory analysis of cultures is unreliable and is unlikely to be available to all study participants as in the Phase 1 setting.

Performance of the statistical model will be determined by calculating its sensitivity, specificity, positive and negative predictive values comparing calculated risk with clinical diagnosis of sepsis occurring within 72 hours of birth.

The investigators anticipate Phase 3 will take place between late 2024 and March 2025.

The TRIPOD (Transparent reporting of a multivariable prediction model for individual prognosis or diagnosis) statement have been followed in the design of this study and will be followed in its reporting(42). Adjustments will be made as necessary with the release of upcoming artificial intelligence-specific guidelines.