Trauma Follow-Up Prediction (Project 2: Aim 2)

The technological convergence of mHealth and machine learning provides an unprecedented opportunity to transform injury care in SSA, particularly for disadvantaged populations. The ubiquity of mobile phones and the advent of mHealth provides a novel opportunity to improve injury care in SSA. Given high levels of mobile phone penetration in Cameroon (85% to 95%) and elsewhere in SSA, the investigators designed and piloted an mHealth, phone-based 7-item screening tool for trauma patients to predict the need for in-person follow-up care after discharge. If effective, this approach could efficiently identify the subset of patients most likely to benefit from follow-up care, which is more feasible, scalable, and cost-effective than blanket advice for post-discharge care. The investigators found that phone follow-up is feasible and acceptable and a validation study revealed good correlation of the screening tool with an independent, in-person exam.

Investigators will build upon their prior research and use data science to improve, implement and evaluate the mHealth screening tool, with the ultimate objective of reducing the crippling burden of injury. This will be achieved by leveraging on machine learning, which has demonstrated promise in optimizing trauma care and trauma systems.The novel combination of mHealth and machine learning provides a powerful opportunity to transform access to health care for those least likely to receive it. Building on existing knowledge, the investigators hypothesize that a data-adaptive, machine-learning approach to outcomes prediction could radically improve survival and reduce morbidity after injury in SSA.

Investigators will apply a machine learning approach to adaptively optimize the mHealth triage tool, improving the phone call timing and algorithm that predicts the need for follow-up care via a cluster randomized controlled trial. This will be achieved using SuperLearner for prediction and cross-validated targeted maximum likelihood estimation (CV-TMLE) for variable importance, using the trauma registry, contact attempt, and screening survey data collected in Aim 1. The overall goal is to improve the mHealth tool's prediction of vulnerable patients needing follow-up care after discharge. This study shall be conducted over an 12-month period; enrollment in 6 months and follow-up participants for 6 months. Investigators will evaluate the impact of the optimized approach in a randomized study in 10 hospitals with 852 injury patients with the primary outcome of the Glasgow Outcomes Scale-Extended (GOSE) score at 3 months.

We originally planned on using the combination of our mHealth surveys and trauma registry data to select a group of patients at greater risk of poor outcomes, for whom a more intensive follow-up could be beneficial. We also planned on using preliminary data to design a new "intensive" intervention that could be randomized to half of this group of selected patients and the other half to the standard of care mHealth survey (conducting surveys at 2 weeks, 1,3, and 6 months). However, we have discovered that 1) there was significant variability in the ability to contact patients, and given issues related to cellphone use and other factors, it was unclear that more intensive contact with this patient population is feasible, and 2) this variability allowed us enough natural experimentation to carry out our goals of this aim, which is to look at the intensity of contact by health workers and its impact on future health outcomes (GOSE score, death, etc.). This latter point means we can use the proposed Targeted Learning methodology with the mHealth data to develop new rules for engaging with patients after discharge, without the need for a sub-study with a randomized component. For instance, we have measures of the intensity of contact at 2 weeks, which now include patients that 1) never were reached for this wave, 2) were reached but did not start the survey portion, 3) they started but did not complete survey, and 4) both started and completed survey. We will then estimate (using our expertise in machine learning for causal inference) the association of this variation in contact and how that relates to future health outcomes (e.g., GOSE score collected at the 1 or 3-month survey). Initially, we anticipated that the four surveys for nearly all patients would be characterized by 4) above. However, we have experienced enough variation in the contact intensity (e.g., patients in categories 1-3 above), that we can identify both the causal impact of contact intensity (adjusting for a considerable number of baseline confounders in the CTR data), can use estimation of the CATE function to determine which patients are predicted to benefit the most from contact and survey, which patients are predicted to be harder to reach at a particular survey, and other functions that can provide actionable information on developing a protocol for patient follow-up.

This allows us, by using the natural variation in the mHealth surveys across and within patients, to accomplish the goals of Aim 2, without a new randomized sub-study. It will also require little to no change in the proposed methodology by which we identify patients at most significant risk of poor outcomes, who we predict will be hard to contact, and those patients that will benefit the most from being contacted by the mHealth surveyors.