ANALYTICAL EYE/I

In this section we wish to describe shifting role boundaries in a startup. As a Senior Data Scientist my role included tasks across disease classification, deep learning research, business development and data wrangling for model training. I often occupied a ‘wear-all-hats’ position in which there were more tasks than people and more roles than could be filled.

Trained both as data scientist and anthropologist the borders of my skills have never been settled. This meant that my ways of thinking and valuing people’s insights were not fixed in terms of algorithm development. On one level psychically (a debate in my head) and outwardly (a debate with colleagues and product agendas), a kind of slippery battle waged across domains of knowledge that were valued among one group and not valued among another group. The value I placed on knowledge and expertise was always open for revision depending on the tribe I was working with. This led to occupying various roles even when my title stated a very bounded one. Carrying out various tasks, often contradictory ones within certain roles has been a professional tension that has persisted. For example, I often went from data scientist to product and partnership development taking back from partnership meetings client suggestions that were not recognized until only much later or too late. I caught myself speaking ethnographically instead of with the brevity of an engineer with a definable problem and outcome. Observational thinking was occasionally valued, but often not, and sometimes even scorned. The anthropologist’s seeing with an eye towards levels of knowledge and a data scientist that must get “shit done” did not always rest easy together. Ethnographically stepping back and technically stepping into highly productive algorithm development was exciting but it could also be exhausting. One had to be elastic and thick skinned.

Authenticity was a sticky matter. I had to have constantly revisionary set of understandings on how data science, anthropology and machine learning specifically worked together and for whom. I lived and worked among revisionary sets of ideas that were never wholly laid out and never fully given over to my colleagues. Being authentic as a trained data scientist, deep learning researcher and anthropologist was not, however, a negative affair, it was one that had no user manual and no clear-cut boundaries. It did require summoning creative parts in oneself. It was a skeptical stance towards algorithms, and an enthusiastic stance towards their possibility.

Role boundaries and formal positions within our start-up could crash into product sprints, team meetings, and patients and physicians we were supposed to serve. This was our company’s way of operating professionally. It was a specific way of working and acting upon knowledge that was cherished and knowledge that was left opaque. This was my direct experience of the organizing potentiality and growth of data science thinking, and not just its application.

THE CONCEPT OF THE “PATIENT”

One other role, the role of the patient, also eluded the makeup of our early start-up.

Patients as abstract beneficiaries, radiologists as domain experts, hospital administrators as users shaped our evolving and tenuous company culture. An embodied sense of what a patient was or who such a patient could be was often abstracted and would appear lost to busy organizational members. However, the patients that were to be served and the radiologists whose diagnostic skills were to be enhanced by algorithms were not really lost but instead were not shared among team members. ‘One-on-one’ walks could not get at shared understandings and would veer off into my or your medical story. There were many “patients” many “radiologists” many ideas of how algorithms would “enhance” the diagnostic encounter. The patient was always elusive and suggested a different kind of ethnography in order to render up these clinical-technical nuances that were embedded in such a machine learning organization, algorithm development and ideals and organizational members under intense pressure to build, scale and commercialize.

The problem of getting the patient in view was mixed with the problem of getting our models in view for productization. We remained far off from an actual product. We struggled with what kind of patients would benefit from our work, were these younger/older patients, smokers/non-smokers, male/female, those who only could afford sophisticated algorithms applied to their chest CTs or MRIs, were they local or in the developing world, were they near-term or years away from benefiting from our work?

A team member brought this confusion into relief when he read from a patient letter he had received. He was the one practicing MD on the team. The patient’s letter was heartfelt and delivered to his home and he brought it in that day to bring through the door an embodied sense of a patient who had undergone both, good and poor diagnostic experience. He had been his patient. This is a paraphrase of the letter.

Dear ______

I wish to thank you for what you have done for me and my family. I’m not sure if you remember me but I was the difficult one who kept asking questions and you were the one doctor I could rely on during my time that tried to answer them. My wife said you were in and out most of the time but I felt you were at my side. I’m not sure how you found the tumor when others missed it, I guess I don’t know how doctors cannot see things that can kill you and someone like you can see it. I really don’t understand medicine but I know you were there and helped us and I’m grateful to you.

I’m delivering this to your home because I wanted to make that effort to come to you as you must have come to me at my bed[side]. I think you saw me as a person not as another patient, I want to believe this and will hold onto it.

Again I and my family wish you much happiness and success in the future and again thank you for your professionalism and care.

Sincerely,

____________

This was delivered in a company-wide meeting. It was an intervention of sorts in an attempt to ground our efforts in an ethics of patient care and an embodied picture of an actual patient who was spared by good diagnostic care and who was grateful and alive to attest to their care. It was also a way to open up a discussion on the consequences of diagnostic error which our algorithms were to correct and reduce. Questions were asked about this patient but then we moved abruptly into an investor meeting. This quick move away from the reality of this letter was indicative of knowledge and experience that could not be taken in or absorbed. A few months later doctor colleague brought up this incident with a sense of shock that any notion of a patient just floated and could not be grounded in our push to build diagnostic algorithms that would help to save such person’s lives. Between the care of algorithm development and diagnostic care of patients a huge gulf existed. The perceptual limitations of algorithm builders were well on display.

Upon reflection the awkwardness of the team’s responses or non-responses to this patient’s letter was of a different order. It conjured up many patients, many diseases, and many possible projects that brought out the ambition in team members to assist doctors in avoiding missing a critical diagnosis. The letter paradoxically, paralyzed and mobilized the team. Our team was not trained to take on and manage patient suffering, they were not trained to hold threats of mortality or the threat that mortality may come anytime around the corner by a missed diagnosis of an aneurysm or a missed adenocarcinoma (most common form of lung cancer). The team was not emotionally capable to both, take on the threat of diagnostic error, or to accept the overwhelming gratitude that comes from having another chance at life. The work of algorithm development and the work of giving more years to a person’s life were at professional and experiential odds.

As a data scientist the teams I have worked with have fallen along a spectrum of well-integrated to chaotic. Our team was well-integrated at times and fell apart into chaos at others. Had I been an ethnographer coming in and out of the organization these moments may have been missed and I would have come away with a completely different picture of team dynamics. For example, after this letter was delivered we floated an idea of a “patient committee” inclusive of patients who had experienced lung cancer and who perhaps had a missed or under diagnosis. Everyone had a different idea of what a patient was, who a patient was and what kind of patients we could recruit. The topic of patient insights being integrated into algorithm development eventually went by the wayside and were reduced to domain knowledge, sent to the periphery of the organization and put on the shelf as patient consultants, when needed. The conception of the patient was marginalized even when it was emotionally charged and required for a deeper understanding of doctor-patient interaction, diagnostic product workflow development and understanding of a medical professional’s insight and error. Even when we clearly needed to hear the voices of patients as part of our R&D, somehow we could not accommodate those voices.

Our operations manager put it well: “we all have been patients but we can’t imagine their needs when we’re here. It’s like we stop thinking of ourselves as whole people, here we are only parts of our selves. It’s strange, my father died of lung cancer, you have extensive experience in the field but the team has a hard time mapping these experiences to their work.” I asked her what she was really getting at.

“It’s hard to build what we’re building [algorithms] on pain and suffering of others.”

I reminded her that we were building algorithms to avoid such suffering through enhancing radiologist perception, to save lives. The patient’s letter was a “success story.” She shook her head as if to say, that was only part of the story when we were outside the organization, on retreat, but not accepted inside the organization where the “real” work got done. Emotions and patients with all their suffering and pain were messy, algorithms were to protect us from that messiness.