Optimizing palliative care: When algorithms predict a patient’s death

“You only have a few months to live.” That sentence is undoubtedly something no one ever wants to hear. The bearer of such a message is usually a doctor, a person who knows about illness – and whether a patient will survive it or not.

According to the 2016 report on nursing care by German health insurer DAK, 60 percent of Germans say they would like to die at home. The most common reason given is that dying in one’s accustomed surroundings would be easier and more dignified. In reality, however, things are much different: 75 percent of Germans die in a hospital or nursing home.

Since 2007 patients in Germany have had a legal right  to “specialized home-based palliative care” if their condition cannot be treated and is considered terminal.  In such cases, the patient remains in their own four walls, attended to by an interdisciplinary team of doctors, nurses, psychologists and counselors who are specially trained in end-of-life care.

There are a number of factors, however, that can make it difficult for patients to access such care:

  • There are currently not enough resources for providing home-based palliative care throughout the country. In many regions, especially rural ones, few specialized teams exist, which is why many seriously ill patients do not receive at-home palliative care, or it is provided later than it ideally should be.
  • Too few doctors and nurses in Germany are currently trained to provide at-home end-of-life care.
  • In some cases, physicians hesitate to prescribe such care even though patients with serious illnesses would benefit from it. This is because doctors can sometimes be overly optimistic, believing, for example, that the patient might recover. Or because they simply don’t have enough time.
  • Making a proactive decision far in advance about which patients might benefit from palliative care is a complex process, especially for practitioners who lack training.

But how does a physician decide which patients would benefit from at-home palliative care? And how do the possibilities change when it is not a single practitioner making the decision, but when they are assisted by a computer program?

Algorithmic systems are already being used, and further developed, to provide medical care – a trend that can hardly be reversed. There are already systems in place calculating the chances of a lethal event taking place – when hospital patients are connected to monitors tracking their vital signs, for example. Medical professionals and data scientists have even been working on a way to predict the probability that a terminally ill patient will die within a given amount of time. Two examples from the United States are currently getting a lot of attention:

1. Algorithm developed by Aspire Health, a new player in the area of palliative care

Aspire Health, which has received funding from Google’s parent company, Alphabet, developed an algorithm that can predict when people with serious illnesses will die. “We can say which patients will die in one week, in six weeks or in a year,” said Bill Frist, co-founder of the US-based firm, to the Wall Street Journal in 2016. The algorithm is used for seriously ill individuals, e.g. those with late-stage cancer or a terminal chronic condition.

The company’s goal is to use the algorithm to prevent patients from undergoing unnecessary treatment. Instead, they should receive professional palliative care in their own home instead – and sooner than they otherwise might. The Aspire program examines data from patients’ medical records and from the outcomes of frequently used therapies to identify who would benefit from palliative treatment.

2. Algorithm developed by Stanford University

Researchers in the Department of Computer Science at Stanford University have taken a similar approach. In a current paper, the researchers describe a neural network that they trained using health records to predict whether hospital patients will die within 3 to 12 months.

As the Stanford team led by Anand Avati states in the paper: “Our approach uses deep learning to screen patients admitted to the hospital to identify those who are most likely to have palliative care needs.” These are people, the paper notes, which conventional care systems might potentially overlook. The predictions are meant to help doctors meet patients’ needs at the end of life, especially by getting palliative-care specialists involved early on.

Commercial use vs. public interest

The timeframe of 3 to 12 months is particularly significant in palliative medicine, since that is when end-of-life care benefits patients most. The systems developed by Aspire Health and the Stanford researchers are designed to achieve the same goal: identify patients who are most likely to die within this timeframe at an early stage of treatment. That would make it possible to provide them with a high level of at-home palliative care sooner than patients generally receive it today – thus helping them die peacefully and in a dignified manner.

In contrast to the Stanford researchers, Aspire Health is running a commercial business. The company’s main service is providing various types of at-home palliative care. Interdisciplinary teams of doctors, nurses and psychologists attend to seriously ill patients in the patient’s own home while also supporting family members. The company focuses on providing relief from pain and other symptoms. It also offers a telephone hotline that patients or family members can call 24/7.

Since 2013, investors have committed more than $50 million in risk capital to the former startup based in Nashville, Tennessee. Aspire Health now operates in numerous cities in 25 states and in the District of Columbia. Bill Frist, the company’s co-founder and the chair of its Board of Directors, represented Tennessee for 12 years in the US Senate. He served as Senate Majority Leader for four years and was even considered a possible candidate for president in 2008. A former heart and lung transplant surgeon, the Republican politician is well versed in the policy framework impacting the US health-care system and what the country needs in terms of nursing care.

In addition to patient care, Frist has set a clear goal for Aspire Health: reducing the cost of treatment for seriously and chronically ill patients by offsetting it against life expectancies. The idea is to prevent patients who only have a short while to live from having to undergo unpleasant treatments or take unneeded medications with harmful side effects; likewise, patients are spared hospital stays that would do little to improve their condition. What the company promises instead is palliative care that is more professional and cost effective and that allows people to die in a more dignified manner in the comfort of their own home.

Aspire Health’s algorithm: Big promises – but does it work?

According to a study by the US-based Henry J. Kaiser Family Foundation, about one-quarter of all expenditures by Medicare, the national US statutory health plan for senior citizens, is spent on beneficiaries in the last year of their life. It is money that is not being spent “smartly, well, efficiently or productively,” Frist says. In contrast, insurers save thousands of dollars per patient when unnecessary treatments are avoided thanks to services provided by Aspire Health. Moreover, the former Congressman says, these services result in exceptionally high levels of patient and family satisfaction while reducing hospitalizations by over 50 percent.

What remains unclear, however, is which treatments Aspire Health considers “unnecessary.” For example, is chemotherapy unnecessary if it will only extend a person’s life by a few months or weeks? Some patients decide they do not want to suffer the unpleasant side effects and prefer to have more quality of life instead. Others, however, want to try every possibility and therefore choose chemo, even though their doctor advises against it – a decision they have every right to make.

The problem is that Aspire Health has yet to demonstrate that its algorithm is in fact effective. There are no public studies available about how high the cost-savings actually are, whether using the algorithm to identify patients actually improves access to care, or how good the algorithm is at identifying the “right” patients. No one outside the company knows how the “death algorithm,” as it was recently deemed by German broadcaster ARD, actually works. It’s a trade secret.

Stanford algorithm shown to be highly accurate

That is the most glaring difference between Aspire Health’s algorithm and the one developed by the Stanford researchers, which can be scrutinized by anyone. The first study provided an initial proof of concept: The patient information used for developing the algorithm was taken from a data pool encompassing some 2 million adults and children who were cared for at either the Stanford Hospital or the Lucile Packard Children’s Hospital between 1995 and 2015.

The data were taken from patients’ electronic records and included diagnoses, completed treatments, the number and type of visual diagnostics (x-rays, ultrasounds, CTs, etc.), prescription medications and the length of hospital stays. The team took the dataset and filtered out some 160,000 cases, which they used to train their model, a deep neural network. They then tested the system using data from an additional 40,000 patients to see how well it had learned to calculate the date of death.

The result: Nine out of ten patients whose deaths were predicted to occur within 3 to 12 months did die within that time. Thus, there were few false alarms. And 95 percent of the patients who were assigned a low probability of dying survived for more than a year.

Human bias: optimism

As research has repeatedly shown, doctors are less adept at predicting a patient’s impending death. An overview study from London, for example, comes to the conclusion that the accuracy of clinical estimates can range from 23 to 76 percent. And multiple studies have shown that doctors often overestimate a patient’s chance of survival – since they are only human and thus tend to be optimistic. Their over-optimism, however, means that patients might receive palliative care later than they should and thus might not get the support they need when they need it.

At the same time, there are many tools and scoring systems clinicians can use to more objectively assess what a seriously ill patient’s chances of survival are. Unfortunately, they do not result in timely and dependable predictions of when the patient is likely to die.

That is why in most cases doctors rely on their own experience. For example, they know what the statistical survival rates are for pancreatic cancer and how long, on average, people have who develop the disease. Moreover, they know the patient’s general condition. And if they have been involved in the case for a while, they know what the patient’s mental state is: Does the patient still want to live? Or are they discouraged or despondent and have come to accept that they are dying?

In Germany, there are no standardized criteria for prescribing at-home palliative care. That means when doctors recommend such treatment for a patient, the decision is based on a range of factors, many of them subjective – i.e. human.

The paper published by the Stanford academics has yet to be subjected to a peer review, meaning by other researchers not involved in the study. And their algorithm has yet to be tested under real-world conditions, meaning with data from patients who are still alive. However, in contrast to the one used by Aspire Health, the Stanford algorithm is not a trade secret.

Lessons for using algorithms within palliative care

Using the Stanford research as a starting point, there is nothing to prevent a public discussion from taking place of “death algorithms” and the ethics underlying them. Such a discussion is greatly needed, since many challenges must be overcome before self-learning algorithms of this sort can be used as a standard part of palliative care:

  • How can such algorithms be checked? The results of machine learning are not predetermined, which means they cannot be foretold in advance and are thus difficult to verify. How can criteria therefore be defined to ascertain whether a self-learning algorithm has passed the test or not? Which requirements must be met before a “death algorithm” can be deemed reliably accurate?
  • A doctor makes the final decision on which treatment a patient will receive. But what happens if the attending physician does not realize that the algorithm has possibly produced a wrong recommendation? How can we empower doctors to better understand the system’s output and recognize mistakes?
  • Who decides how the algorithm will be programmed? Which goal is the algorithm meant to achieve? Which limits should be set on the algorithm’s sensitivity and specificity?
  • Which dataset can, should or must be used to develop an algorithm? Would data from electronic health records suffice? Or should other information be included, such as doctor’s notes and even patients’ statements? Should it be possible to use such algorithms for commercial purposes?

If an algorithm can accurately predict whether and at what point a patient would benefit from palliative care, then it could indeed support doctors in doing their jobs. The more confident doctors are about the prognosis they make, the surer they can be that the treatment they recommend will help the patient as it should.

Many people might be very uncomfortable with the idea of a computer program predicting the probability that someone will die. Yet whether or not an algorithm is involved, doctors are ultimately responsible for the message they give their patients – and for working with them to decide if palliative treatment is the right choice or not.

The prerequisite for this, however, is transparency: Doctors must know how an algorithm works and which data are being used to generate predictions, taking into account related characteristics such as specificity and sensitivity. In other words, how often does the algorithm correctly identify those people who do in fact die within a certain timeframe, and those who have a high probability of surviving beyond a given point in time?

It is the job of the regulatory authorities to ensure this level of transparency. Developers must be held responsible for their algorithms and required to provide proof of their algorithms’ efficiency, effectiveness, residual risk and accuracy, among other factors. Conversely, regulators must develop new benchmarks that reflect these challenges – without hampering the developers’ ability to innovate.

A black box like the algorithm used by Aspire Health leaves too many questions unanswered. After all, the company is earning money with its palliative services. It would be all too easy to assert that its software has been trained to identify as many patients as possible who require its services – leading them to abandon curative treatment too soon. Would we really want to trust such an algorithm?

 

This is the second part in a three-part series on using algorithms to predict death.

Part 1 is available here: “How algorithms can save people from an early death

Part 3  is available here: “Using algorithms to predict death: Lessons learned”



Comment