Elythea: AI to Prevent Maternal Mortality in LMICs

Postpartum hemorrhage is the #1 killer of mothers globally. The CDC estimated that 75+% of these mortalities were preventable if there were better predictive models allowing for earlier preparation/clinical coordination. The second leading cause of preventable maternal mortality is that moms themselves are not aware of the red-flag symptoms to consider if they're high risk and often do not go to the hospital until it is too late. In low-resource settings where provider bandwidth is limited and there is a high barrier for rural moms to travel to the hospital unless it is "worth it", there is a strong desire to solely identify and inform high-risk moms earlier on.

If patients who need a blood transfusion don't have blood prepped, the doctors need to rush to get blood during their delivery, which can take 10-30 minutes (up to 2-6+ hours in rural/developing regions), risking death. In African countries, you need to have a family member donate a unit of blood before you can receive one, which adds extra time when a mother is hemorrhaging. These hospitals lack the resources to assess blood type/have blood prepared for everyone but do have the capacity to prepare blood for the few high-risk patients. By predicting months in advance, at the point-of-care, we give doctors/moms months to get type/cross-matched blood ready before labor even happens.

When we interviewed 80+ obstetric providers and asked them what complication they were most worried about/wanted to prevent, >90% pointed us to Postpartum hemorrhage and hypertensive eclampsia. >97% of providers reported being able to do something in advance if they knew patients were at risk, affirming that earlier knowledge of at-risk patients would tangibly help change their clinical management and improve their odds of saving their patient's life. 

Doctors get burnt out spending extra time on preventable complications and extending patient backlogs, hospital systems lose millions in funding from having higher mortality rates/poor resource allocation, and insurance companies pay $10B extra (US alone) on preventable complications. In African countries, government systems (specifically the Ministry of health) lose $1B+/year paying for these costly complications. 

Globally, 140M births occur, where >80% of patients have a mobile device/are patients at a clinic with a mobile device capable of using our platform. 14M moms have PPH annually, causing 70,000 maternal deaths globally, >60% of which are preventable (WHO). Up to 11M moms have eclampsia, globally causing 50,000 maternal deaths each year. >25M moms have an unexpected, emergency c-section each year.

>99% of maternal deaths occur in developing countries, yet there are no widely adopted prediction models and most facilities rely purely on clinical judgment.

African obstetricians currently have medications (oxytocin/TXA) and intervention mechanisms (tighter follow-up, active management of labor, specialized maneuvers/MFM attendings) to prevent/treat these complications, but just don't have enough resources/time to allocate for every patient--just the high-risk ones. 

But, doctors in developing African countries have as low as ~1-3% accuracy rates (NIH, Reproductive Health) in detecting hemorrhage, and developed hospitals have red-flag rubric point-scoring systems that have <50% accuracy rates. 

Elythea is a machine learning mobile/web app platform that obstetricians in any setting use to predict risk for pregnancy complications (postpartum, hemorrhage, eclampsia, preterm-labor, etc) before they happen. We use ML models that are better at dealing with multidimensional medical data than linear models/red-flag checklist systems.

We use sociodemographic/clinical history information that is available at the point-of-care, so we can make the prediction as early as the 1st trimester, allowing doctors months to prepare in advance! Our platform can intake manual input of patient features in an easy-to-use mobile question flow which is especially helpful for nurses/providers in low-resource regions. Every doctor we interviewed across 5 African countries affirmed that a mobile app (1-time download through mobile-data) would be easily accessible for them and their care staff. NCBI publications substantiate this, supporting that >90% of Sub-Saharan African doctors owned a mobile device. 

By predicting adverse outcomes, we allow doctors to allocate their finite resources (medicine, time, specialized consult) for the patients that truly need it. For moms at risk of hemorrhaging, doctors can coordinate family members to donate blood months before delivery and have blood-type-matched blood ready before the mom even goes into labor. During/prior to labor, oxytocin can be prophylactically administered, along with tighter clinical followup. Moms who are aware they are high risk can get clinical consults on specific red-flag symptoms to watch out for, can have diet/exercise regimen modifications, and are more likely to travel to hospitals when they detect concerning symptoms. 

No existing models predicting PPH have been prospectively tested in rural/African cohorts, despite 99% of PPH deaths occurring in low-resource regions like Sub-Saharan Africa.

We prospectively tested our models on 2,000+ women enrolled from 13 sites across Cameroon, Nigeria, Uganda, Texas, and Rhode Island in predicting complications like PPH, c-section, and eclampsia. We have demonstrated a higher accuracy, F1, and AUC ROC metrics than all other externally validated gold-standard models/judgments used in clinical practice. 

Of existing models, we address the following limitations of the status quo: 

1) Current models are trained on low number of patients (<10-50k): 

We trained our models on >10M patients

2) Current models have a low number of "positive samples" to train their models on. ~1% of patients in the data need a transfusion, which is really bad for training models, causing poor true positive detection rates:

We used proprietary models, hyperparameters, data augmentation/feature engineering techniques, and weight scaling algorithms, which took ~2 years to develop and are state-of-the-art and utility patent-pending.

3) Current models can't be used until later in the pregnancy (when it's too late/the mother has already developed complications of pregnancy) and require expensive/invasive blood tests: 

Our models can be used *at the point of care*. We primarily use clinical history/demographic information (ex: number of previous pregnancies, smoking status, education-level, age, etc), most of which is available and can be used for prediction as early as the first trimester! This makes it accessible to developing countries and low-resource rural regions. 

Elythea demo:

We serve moms in developing countries and rural regions. Currently, these moms face the highest complication burden and constitute 99% of the global maternal mortality. 

These moms come from widely different backgrounds -- ranging all the way from unplanned teenage moms to moms of advanced age, they span different education levels, different income statuses, and different levels of familial/spousal support. 

Yet, they all share multiple unfortunate and inequitable burdens -- they typically live far away from healthcare facilities. It is a prohibitive barrier to travel all the way to the nearest pregnancy care facility. There are typically local clinics with midwives that staff these clinics, but moms that have complications (like hemorrhage or eclampsia) must be transported (while they are complicating) to advanced facilities (typically tertiary care clinics closer to the main cities). 

Because they are unaware of their risk status, many high-risk moms ignore red-flag symptoms (that they have not been counseled to watch out for) because of the potential financial and logistical burden that traveling to the nearest hospital imposes. Many high-risk moms also go to their local community clinics expecting a normal pregnancy, and have precious hours wasted while they suffer through life-threatening complications like hemorrhaging or eclampsia as they are transported to a tertiary care facility with the resources/doctors to actually take care of them. 

Moms at risk of hypertensive disorders like eclampsia have defined prophylactic antihypertensive medication, exercise regimen, and diet modification,  that they can undergo to drastically reduce their risk of their complication (by 80%!). In potentially life-threatening cases, earlier induction of labor can be scheduled to optimize the likelihood of the mom and baby surviving. Currently, because doctors have poor accuracy in predicting these complications, these moms have to wait until they have their complications and must wait to be treated, risking death.

Elythea provides promise for these moms to live through their complications (and even prevent their complications in many cases). Given that >80% of Africans own mobile phones (UN 2016), moms in these regions are able to input their basic clinical history and demographic information to see if they are at risk of complications. This gives moms a tangible reason to get advanced medical consult, initiate prophylactic diet/exercise/medication, schedule deliveries at better-resourced tertiary care facilities, and allows moms to get better clinical consulting, allowing them to watch out for red-flag symtoms better.

Doctors/midwives in these low resource regions (>95% of which own a mobile device), will be able to prescribe better treatment, have improved clinical coordination, and can know which high risk patients to schedule their limited experienced attendings for.

My mother had a bleeding complication following a cancer-related hysterectomy--this opened my eyes to the field of hemorrhage and prompted us to conduct our user interviews with 80 obstetric providers across 5 countries (4 of which were African countries). We spoke to doctors that worked in different socioeconomic strata of Africa -- ranging from doctors in the developed capital of Lagos, Nigera to doctors working in rural Uganda. We made a sincere effort to speak to as many African providers as possible to understand what complications were their biggest concern, where the true pain lay, what resource limitations they were working with, what the most important criteria in a predictive model would be, and what cultural considerations to integrate. 

Melissa Bime is our key partner and trusted advisor. She is a former Cameroonian nurse who grew up in Cameroon herself, being intimately acquainted with the African healthcare ecosystem and its structural/cultural environment. She founded a YC-backed clinical research startup initially focusing on blood transfusions/blood bank infrastructure working with African hospitals (gaining deep knowledge of African medical systems). We were able to leverage her existing network to conduct our African trials, coordinating patient recruitment/ethics, and PI recruitment. 

Moreover, my (Reetam) professional background led me here: 8 years ago, I founded of global biology education nonprofit, Junior Medical Academy, which has since been accessible to 5,000+ students across 30+ countries (primarily developing African countries). With this nonprofit, I developed a novel model to reach non-English speaking students in rural African regions previously unreachable by Western nonprofits through a grassroots movement, by working with student communities in the capital to reach out and teach to students in rural villages. I have deep experience working with local African communities, expanding to rural African regions, and have been able to leverage my existing network for clinical trials.

Our MVP platform, use time, variables included, UI, etc have all been screened across the dozens of African providers that we have been able to interview. We have incorporated cultural/logistical changes they have requested and are making an active effort to work with these communities. 

We just funded a randomized control trial spanning >1,500 moms and dozens of African providers across rural Uganda, Nigeria, Ghana and Cameroon. We are recruiting currently, >100 African moms have been through our platform. We have active boots on the ground and close relationships with African obstetric PI's who are excited and willing to help inform the implementation of our platform. 

We are also working with rural clinics in Bihar to deliver our predictive models to screen at-risk pregnant patients. We have signed contracts to reach up to 500k pregnant moms in rural Bihar and are working with the rural patients, doctors, and clinic operators to develop and integrate our solution, ground up and fully informed by the people directly impacted by our platform. 

We have prospectively tested our models on data from 240 moms across 10+ sites in Nigeria and Cameroon. 

We have funded a randomized control trial pilot for >1,500 rural Ugandan, Nigerian, Cameroonian, and Ghanian moms over the next year, >80 moms have had Elythea impact and inform their care journeys. 

In General: 

The majority of locations (>70%) use clinical judgment and developed settings use high-low rubric scores with <50% accuracy and 0.52 AUC. We are able to predict hemorrhage with 78% accuracy and 3x higher sensitivity. 

Competitor Companies are in 2 Buckets:

1) Physical Devices:

Ex: NUVO and the Oli Device. 

They use a physical monitoring product moms must wear to record pregnancy data and try to predict complications based on physiologic data. These are expensive, inaccessible in low-resource/rural regions, and require long-time use/physical wearing of device. 

2) Blood-Based Biomarkers:

Ex: Mirvie and Sera Prognostics.

 Use RNA-sequencing data to predict major complications of pregnancy--it is extremely useful but requires invasive testing, is expensive, takes 1-2 weeks to get results, and is completely inaccessible and unaffordable to rural regions where >95% of the hemorrhage/eclampsia burden lies.

Here's why that's bad:

Patients are reluctant to use invasive tests and pay the high bills; there is very low clinical utility to predict at late pregnancy stages (at which point it is too late for the doctors to clinically intervene/prevent complications, especially in low-resource settings). 

Here's how we address that:

We can make our predictions as early as the first trimester, only take 60 seconds to use, can be used in any location with mobile phones (>85% of patients in developing countries), and require NO physical/blood tests. Our SaaS model lets us scale rapidly without having to process blood draws, and lets us give risk scores in <1 second versus 1-4 weeks like competitors. 

Any venture using AI will hit the class imbalance issue: low positives/low proportion of minority patients will lead to poor AUC/accuracy. If only 1% of your patients bleed, your models will erroneously predict "Healthy" for most patients. If <10% of your training data is from black moms, your models will fundamentally be biased against minorities, which is unacceptable from a lens of health equity. 

Our propietary models/weights/hyperparameters/data-augmentations and our hyperparameter weight-scaling algorithms are our advantage. By being able to expand the model's distribution, we drastically improve how our model predicts for hemorrhaging moms and minorities.

Here's how we will change the market:

Right now, there isn't a huge emphasis on "point-of-care" predictive models. Most need late-stage variables that are inaccessible. We want to change that. As we gain market adoption, we want to sway the market to follow suit and make their diagnostic/predictive devices emphasize earlier detection/diagnosis to actually allow doctors/patients to take measures to intervene/prevent complications. 

We also want to push the predictive tech market to use generative AI to augment models. As medical care gets increasingly advanced, we will hit a “diminishing positives” issue, where lower rates of serious complications make it increasingly difficult for models to predict these complications. We want to pave the way for all models to implement algorithms assigning higher weighting to low-representation outcome variables and minority patient data, truly and algorithmically promoting health equity. 

1-Year Goals:

Within 1 year we want to conclude our clinical trials, having hard, statistically valid evidence demonstrating the statistically significant decrease in costs, death, and maternal morbidity from Elythea usage.

We have already finished up prospective trials demonstrating that Elythea models have higher accuracy, AUC ROC, and recall metrics than existing methods and have funded and submitted ethics clearance for our randomized control trial. We already have established partnerships with the nurses and PI's conducting the trials, and have the infrastructure in place once we get approval within the next month. 

We have signed contracts to reach >50,000 moms over the next year across rural India, Southern suburbs in the US, and mutlipel rural African countries. We hope to reach these patients and prevent >10,000 preventable complications and lead to  at least 1,000 maternal mortalities uniquely avoided with Elythea usage.

5-Year Goals:

In 5 years, we have currently signed contracts to reach >500,000 pregnant moms globally, in the past 3 months of commercializing. As we continue to grow, we want to reach >10,000,000 moms across the world, sign on 100 obstetric facilities in the US and 250 obstetric hospitals across African, South American, and Asian countries to use our platform, and accurately prevent >50,000 avoidable maternal mortalities.

 We hope to have our prediction accuracy be >90%+ with the incoming data and advanced deep learning augmentation that our models have been fine-tuned under once we collect hundreds of thousands more data points from African/rural Indian patients. All this collected data will be proprietary and will directly go toward helping us tune our models to perform equitably, and predict optimally for an African patient base. 

Currently, no such openly accessible large-scale databases exist for obstetric African patient cohorts to train ML models upon specifically tracking hemorrhage/eclampsia. Therefore, high-powered US databases, like our existing dataset of 10M+ patients from the CDC the closest proxy currently available to train models that would be potentially generalizable to African patients. We have been able to have proprietary frameworks and weight scaling algorithms to perform better on African cohorts than current US models perform on US patients. But we don't want to stop here. Our vision is to source thousands of African data points and use this data to drive positive change by having models best equipped to detect for the populations of moms that need it the most. We hope to be uniquely situated to be able to do that in 5 years!

Currently, moms that are high-risk don't know that they are high-risk. Thye miss clinical symptoms, hesitate to go to the doctor if they are in rural regions due to financial barriers and don't take indicated precautions that can help prevent their complication. 

By giving them specific risk scores (accessible by a 60-second at-home questionnaire that asks about their demographic and clinical history (like their age, education status, number of children, etc -- things every mom knows off the top of her head), moms will know which complications (like eclampsia or hemorrhage) they are at risk for. This will prompt them to go to the doctor when they would not have otherwise. The UN isolates lack of maternal awareness of high-risk symptoms as one of the leading causes for why the majority of maternal mortalities are preventable. 

 Doctors can use Elythea to find out which moms are high-risk and will know to schedule more followups with them, specifically counsel them on the complication they are at risk for, administer prophylactic medicine to reduce their odds of complications and schedule more experienced attendings for their delivery.

When we interviewed 80 obstetric providers across 5 countries, >97% of providers reported being able to do something in advance if they knew patients were at risk. Publications like the Lancet substantiate that enhanced postpartum hemorrhage care regimen can reduce maternal mortality by >90% and can save hospitals >$1M/year.

Full time: 1, Advisors: 5, contractors: 30

2 years

Our team is diverse -- we span BIPOC, genderqueer/LGBTQIA folks, and native African women who have grown up/worked in the African healthcare system. Our founder's family came from rural India and are personally acquainted with the healthcare challenges and maternal challenges that take place in these regions.

We took conscious action to include African women with healthcare experience to make sure we amplified their voices first and foremost and integrated their suggestions/perspectives deeply into the platform.

We want to include more women who have grown up in rural regions and have personally gone through pregnancy complications like hemorrhage, who can give personal insight into what the problem looks like from the mom's perspective. We also hope to add team members who have worked in digital health startups operating in developing countries and people with a cultural/anthropologic background who bring the business network and diverse perspective to complement our technical team. 

How we promote diversity:

1) Clinical judgment has been shown, time and time again, to have biases against historically underrepresented women, immigrants, and LGBTQIA+ folks. By having objective, ML-predicated models (which fit complex statistical equations to millions of patients' worth of data), we get objective metrics, agnostic to any racial biases, to make predictions off of. This helps to combat racial/systemic prejudices doctors may have. It's no secret that black women are overlooked by medical professionals -- we hope to provide objective means to amplify their voices. 

2) Women in rural/developing countries have a 6-fold higher chance of mortality than women in developing countries when giving birth. This is because they lack the advanced healthcare facilities/technology to be able to predict adverse outcomes. Our technology requires NO lab tests, genomic tests, and doesn't even require the woman to be at a late stage of pregnancy. It can be used at the point of care, anywhere, for cheap. 

3) The biggest reason why current models have biases against minorities/LGBTQIA+ folks is because there is limited training data. If less than 0.5% of your patients are transgender, if only <10% of your patients are black women, etc then it's no wonder why current models will perform poorly for these types of patients. Our generative neural network generates synthetic patient records to provide more training data to racial minorities and LGBTQIA+ folks, which have boosted our models' performances for historically underrepresented populations. 

4) We are working with African nurses, midwives, and doctors who truly and genuinely understand the space. They grew up in African cities themselves, have been working in the healthcare systems for decades, and are the best individuals to understand the intricate cultural dynamics/considerations. As we expand our trials, our team/healthcare pool expands, growing the number of input points we receive to make sure that the platform is best tuned for the people who will benefit most from it. 

tl;dr Business model: we charge hospitals a fixed cost per patient; we contract with the government in developing settings and charge per-patient 

What we based our model on:

We based our model on existing predictive model companies working in the diabetes, GI, gynecologic complication, and postpartum depression space. The business model we are using has deep precedent, and we are adapting the business models of similar companies to suit the appropriate cost burden posed by pregnancy complications. 

We have 2 distribution channels for selling: in the US vs in African/developing countries

In the US:

If you look at the intersection of moms missed by current systems and moms that are caught by Elythea, we are able to catch ~35% of moms missed by the status quo. Of this 35% of moms, ~15% will have preventable/significantly reduced complications with earlier intervention (uniquely prompted by Elythea). 

This 15% of preventable complications translates to a burden of ~$2M/hospital/year. This is due to additional OR beds (which can't be utilized for higher billed procedures), skilled nursing resources, physician time, sanitization equipment costs, etc. Unexpected complications like hemorrhage, eclampsia, and emergency c-section cost up to $28,000 more per mom. There is an aligned payer incentive here between moms, hospitals, and insurance companies. 

Of this $2M/hospital/year burden, we just take a 5% cut (100k) and spread it out of the total number of patients the hospital sees per year, which comes out to ~$50/patient. This is below market rates that hospitals are used to paying -- current predictive models take a 10-15% cut and charge anywhere from $50-500/patient. 

In African/developing countries:

The business model changes slightly in African countries where most moms are covered by public insurance through the government. In these cases, the Ministry of Health pays for the additional costs incurred by pregnancy complications and is the main stakeholder. We will charge the Ministry of Health a per patient charge using a similar financial model to the US hospitals, adjusted for the local costs/patient volumes. 

The other proportion of patients in African countries typically pay out of pocket. For these users, we will charge a small fee per mom directly to the users proportional to their region/income level. (Within the $1-8 vicinity). Although the cost is less per mom, the user base is far larger and completely untapped currently. 

In the past 3 months of commercializing, in the US, we have signed a contract and 2 LOIs spanning a $1.2M ACV spanning 55,000 patients. We also signed a contract in rural India spanning a $1.2M ECV for the pilot (til end of 2024) which automatically converts to $6M contract volume. 

We are an embedded model: the social and business enterprise are the same. As explained above, we will charge a per-patient fee to hospitals (most common) and governments/insurance companies where applicable. We directly target the stakeholder that loses the most money, charge them a small fee per patient, but save them millions in the process. 

We are a software as a service -- in hospitals with an EHR, we can sync into the EHR and provide instantaneous predictions for each patient. In settings without an EHR, we are a mobile app accessible by the majority of healthcare staff and moms. 

We expect the costs to be minimal -- it costs us <$0.10 per patient, while we charge hospitals ~$16 per mom. Our revenue will be enough to sustain us. Furthermore, we aim for our profitability to be an incentive to work with nondilutiveWe hav pitch competitions in the early stages to raise investment capital to accelerate our trials, and once we have proof of profitability, we want to leverage that to gain capital from VC firms. 

Additional go to market strategy:

We have initiated the largest randomized control trial ever conducted in this space across 1,500+ moms across 3 countries to demonstrate a tangible reduction in maternal mortality, costs, and adverse outcomes through publishing our results in a peer reviewed obstetric journal. We would then use these published results as ethos to then sell to hospital administrators and obstetric facilities across the country/world. Peer-reviewed publications are the gold standard way that all currently used models have been scaled and marketed, so our next steps forward have deep precedent.

We will start by converting our clinical trial hospitals, and will directly market to hospital admin from referrals and at OB conferences.  Elythea scales as a software-as-a-service platform. In developed settings, we sell to hospital administration and seamlessly integrate with all EHR systems. In developing settings, we are offered as a mobile app (accessible to >95% of African doctors); distribution through android/IOS app stores. 

Obstetric hospitals in the US get "graded" by the Joint Commission (overarching organization governing hospital funding/rules) based on key criteria like maternal mortality, hemorrhage, c-section rates--if hospitals have poor rates, they get funding taken away and if they have good rates, they receive additional funding. Besides these funding incentives, we will also work with the Joint Commission to target the highest-need hospitals with the worst mortality/hemorrhage rates that would maximally benefit from Elythea. This is strategically beneficial, it helps to maximize early market penetration with the hospitals that would need us the most and provides an incentive for their competitor hospital systems to use Elythea to avoid losing patients/funding. 

We can also work with large health plans in the US to charge on a per-member-per-month basis to help reduce complication rates (and thus reduce costs that the health plan incurs for the patient). 

We will employ a similar strategy in developing countries by working with the governments that lose the most money/pay the most for postpartum hemorrhage and maternal mortality. There is precedent for these government systems mandating the use of certain procedures/tools that have been shown to reduce mortality/costs, which we hope to leverage for Elythea.