Keywords

1 Introduction

The following are recognized by transnational think tank The Millennium Project as significant global challenge (GC) areas: sustainable development, clean water, population and resources, democratization, foresight and decision-making, convergence of information technology (IT), rich-poor gap, health, education and learning, peace and conflict, status of women, transnational organized crime, energy, science and technology, and global ethics [1, p. 17]. This article focuses on health. Specifically, the healthcare industry is inundated with elder, disabled, and partial-use care issues such as falls in homes with no available aid present. This article’s thesis is that bio-support in the form of biomimetic artificial intelligence (AI) is not yet fully-exploited within this problem space.

This article will summarize analyses conducted using Hancock’s “Knowledge-Based Expert System” (KBES) [2] on two publicly-available datasets. The first dataset contains various numeric, health-related data from 400 anonymized patients diagnosed with chronic kidney disease (CKD). The second contains the same kind of data, but for 569 patients diagnosed with either malignant (coded “M”) or benign (coded “B”) forms of breast cancer. Both datasets were retrieved from Kaggle [3, 4].

In the last place, the potential for a “Holacratic” [5] health analytics organization will be explored. Said organization would be akin to Tapscott’s “Global Solutions Network” (GSN) [6], which was defined as a digital group of public or private individuals with the following four features:

  • Diversity in stakeholders who collectively represent at least two of the four pillars of society (government, private sector, civil society, individuals)

  • Multinational or global presence

  • At least partial digitality with respect to its communications tools and platforms

  • Progressive goals related to the creation of “public goods” (p. 18).

It will be shown that the group working on this article (known colloquially as the Sirius Project) successfully addresses both the criteria of being and need for a Holacratic GSN [7, 8].

1.1 Background

In the movie Big Hero 6, “Baymax” is an inflatable robot with a carbon fiber skeleton who serves as a personal healthcare provider companion. Movie codirector Don Hall said of Baymax that he “views the world from one perspective—he just wants to help people” [9]. Baymax is an example of a fantastical biomimetic AI designed specifically to help people solve their health-related problems (Fig. 1).

Fig. 1.
figure 1

Baymax from the movie Big Hero 6.

Full size image

Biomimetic (“life-mimicking”) systems automate functionality by reiterating biological forms and processes (e.g., early airplane designs were attempts to model winged flight in birds). In this way, biological systems not only provide existence proofs—e.g., “heavier-than-air flight is possible”—they also inform attempts to implement functionality by replicating form in automating systems.

Important sub-disciplines within the science of machine learning have developed in this same way, with the most widely-used cognitive architectures lying along relatively short development paths from biologically inspired starting points. Obvious examples are feed-forward neural networks, reinforcement learning machines, genetic algorithms, expert systems, and binocular/binaural sensation. While the application of sophisticated optimization techniques to these systems can blur unscalable aspects of the underlying biological metaphor (airplanes do not flap their wings), what remains can still be informative.

Could biomimetic AI systems aid patients suffering from breast cancer and CKD? Given the popular overlap at present between the fields of artificial intelligence and machine learning (ML), biomimetic AI could draw from the insights enabled by utilizing ML methods. Results from analyses conducted utilizing a ML algorithm are summarized in the following sections.

2 Method

We decided to implement a bias-based Knowledge Based Expert System (KBES) as described by Hancock [6]. The data this expert system was applied to are all publicly available through Kaggle, the popular machine learning repository. The first dataset contains various numeric, health-related data from 400 anonymized patients diagnosed with chronic kidney disease (CKD). The second contains the same kind of data, but for 569 patients diagnosed with either malignant (coded “M”) or benign (coded “B”) forms of breast cancer. Both datasets were retrieved from Kaggle [3, 4].

2.1 Data Preparation

The data preparation process is automated so that it is documented, repeatable, and consistent.

The data preparation parameters for the breast cancer data were:

  • 2 = mode 1 is for training mode 2 is for processing

  • BC_W_T = C45 CSV quantized training file

  • 32 = number of columns in training file

  • 2 = number of Ground Truth classes in training file

  • BC_W_B = C45 CSV quantized blind file

  • 150 = number of reconstruction values for feature quantization

  • 10000 = number of training epochs

  • 0.001 = learning rate

  • 0.5 = proportion of missed vectors to update on

The CKD settings are similar. The CKD data consisted of 400 instances (rows), each having 24 features (columns) and two classes (CKD and no CKD). The breast cancer data consists of 569 instances (rows) each having 30 features (columns) and 2 classes (malignant and benign).

The data for both experiments went through the same conditioning process:

  1. 1.

    Each raw feature column was z-scored so that the units across features would be commensurable.

  2. 2.

    Each feature column was histogram equalized. The feature values in each column were then replaced by their histogram bin numbers.

  3. 3.

    The CKD data (400 instances total) was split into two segments: a “training set” (267 instances) and a “blind set” (133 instances).

  4. 4.

    The breast cancer data (569 instances total) was split into two segments: a training set (379 instances) and a blind set (190 instances) (Fig. 2).

    Fig. 2.
    figure 2

    The quantization process.

    Full size image

The data were then visualized in their native high-dimensional spaces to assess the complexity of the classification process. The breast cancer data are shown below (Fig. 3):

Fig. 3.
figure 3

Visual inspection of the conformed feature data.

Full size image

2.2 Belief Accumulation

Non-graph-based methods for reasoning represent knowledge implicitly as functions (usually a collection of rules with an adjudicator, a regression function, or clauses with a resolution method), and perform reasoning by function evaluation.

2.3 Inferring Knowledge from Data: Machine Learning = “Adaptation + Propagation”

“Adaptation + Propagation” might sound Darwinian, but it is actually LamarckianFootnote 1. Animal brains continue throughout their lives to expand their cognitive power by assimilating their experiences, in an (as of yet) unknown way. In light of this analogy, it is sensible to adopt the following empirical definition of machine learning: Machine learning is the automatic organization of execution history into machine resident structures in such a way that machine performance is improved.

The method used for the work described here is an instance of such machine learning referred to as reinforcement learning. In reinforcement learning, a brain (biological or mechanical) interacts with its environment by performing the tasks that arise. For an animal, these might include eating, mating, avoiding injury, etc., while for an image classification machine, patterns to be recognized would be presented and scored by a training algorithm.

When the decisions made by the system result in positive outcomes, inclinations favoring these decisions are strengthened, making them more likely to be repeated in the future. When the decisions made by the system result in negative outcomes, the corresponding inclinations are weakened, making the unsuccessful actions less likely to be repeated. Over the course of many trials, the system gradually adjusts its inclinations to result in mostly positive outcomes: it learns how to succeed (Fig. 4).

Fig. 4.
figure 4

Supervised learning by feedback and adaption.

Full size image

The Adaptation + Propagation Loop is fast and simple. Firstly, an objective function must be defined that enables the Evaluate and Learn function to tell whether the machine’s last trial is the best seen so far (a trial consists of processing all the instances in the training set). A natural measure of performance might be the proportion of patterns correctly classified. If the machine has improved, the updated weights become the machine’s “New Best”. If not, another set is generated for the next trail, and the process repeated. To generate a new set of weights to try, the current “best set” is perturbed slightly. In this way, previously learned patterns are “propagated” to subsequent trials.

“Bias-based Reasoning” (BBR) is a mathematical method for automating implementation of a reinforcement learning approach in rule-based systems. Reasoning in such a system proceeds by accumulating “evidence” for and against possible conclusions; when a sufficiency threshold is reached, the favored decision is selected. Learning is carried out by making small adjustments to weights inside decision logic rules. During training, the system is run through a large number of decision-making trials. When it produces a correct decision, the weights favoring the decision given that evidence are slightly increased; incorrect decisions result in the corresponding weights being reduced.

Reinforcement learning in rule-based systems enjoys the same advantages human experts derive from this approach; in particular, it supports automated learning, conclusion justification, confidence estimation, and natural means for handling both non-monotonicity and uncertainty.

2.4 Using Facts in Rules

For simplicity and definiteness, the reasoning problem will be described here as the use of evidence to select one or more possible conclusions from a closed, finite list that has been specified a priori (the “Classifier Problem”).

Expert reasoning is based upon facts (colloquially, “interpretations of the collected data”). Facts function both as indicators and contra-indicators for conclusions. Positive facts are those that increase our beliefs in certain conclusions. Negative facts are probably best understood as being exculpatory: they impose constraints upon the space of conclusions, militating against those unlikely to be correct. Facts are salient to the extent that they increase belief in the “truth”, and/or increase “disbelief” in untruth.

A rule is an operator that uses facts to update beliefs by applying biases. In software, rules are often represented as structured constructs such as IF-THEN-ELSE, CASE, or SWITCH statements. We use the IF-THEN-ELSE in what follows.

Rules consist of an antecedent and a multi-part body. The antecedent evaluates a BOOLEAN expression; depending upon the truth-value of the antecedent, different parts of the rule body are executed.

The following is a notional example of a simple rule. It tells us qualitatively how an expert might alter her beliefs about an unknown animal should she determine whether or not it is a land-dwelling omnivore.

figure a

If we have an INCREASE BELIEF function, and a DECREASE BELIEF function (“aggregation functions”, called AGG below), many such rules can be efficiently implemented in a looping structure.

In a data store:

  • Tj(Fi) truth-value of predicate j applied to fact Fi

  • bias(k, j, 1) belief to accrue in conclusion k when predicate j true

  • bias(k, j, 2) disbelief to accrue in conclusion k when predicate j is true

  • bias(k, j, 3) belief to accrue in conclusion k when predicate j false

  • bias(k, j, 4) disbelief to accrue in conclusion k when predicate j is false

Multiple rule execution in a loop:

figure b

This creates a vector B of beliefs \( \left( {b\left( 1 \right), b\left( 2 \right), \ldots , b\left( K \right)} \right) \) for each of the conclusions \( 1, 2, \ldots , K \), and a vector D of disbeliefs \( \left( {d\left( 1 \right), d\left( 2 \right), \ldots , d\left( K \right)} \right) \) for each of the conclusions \( 1, 2, \ldots , K \). These must now be adjudicated for a final decision.

Clearly, the inferential power here is not in the rule structure, but in the “knowledge” held numerically in the biases. As is typical with heuristic reasoners, BBR allows the complete separation of knowledge from the inferencing process. This means that the structure can be retrained, even repurposed to another problem domain, by modifying only data; the inference engine need not be changed. An additional benefit of this separability is that the engine can be maintained openly apart from sensitive data.

Summarizing (thinking again in terms of the Classifier Problem): When a positive belief heuristic fires, it accrues a bias \( \beta > 0 \) that a certain class is the correct answer; when a negative heuristic fires, it accrues a bias \( \delta > 0 \) that a certain class is the correct answer. The combined positive and negative biases for an answer constitute that answer’s belief.

After applying a set of rules to a collection of facts, beliefs and disbeliefs will have been accrued for each possible conclusion (classification decision). This ordered list of beliefs is a “belief vector”. The final decision is made by examining this vector of beliefs, for example, by selecting the class having the largest belief-disbelief.

3 Results

When the trained machine processes a data set, it produces a file containing the classification results for each instance, and also create a confusion matrix so that the specific mix of Type I and Type II Errors can be reviewed in summary form (Fig. 5).

Fig. 5.
figure 5

Results of a 10-fold cross-validation.

Full size image

4 Discussion

Do the findings summarized above speak to the possibility of biomimetic AI systems that could assist elderly sufferers of CKD or breast cancer? Assuming such systems could achieve the 93%+ predictive capabilities demonstrated by the KBES, then the next question becomes of how exactly they would. We envision such a healthcare-providing AI that would serve to entertain and educate a growing elderly population, as well as offer them physical and mental (e.g., affective) support more generally. (Naturally, such an AI system would need to be cost-effective for at least one patient in order to pass the “proof of concept” stage of the production cycle.) A final question posed here is of whether biomimetic AI would be ideal (compared with more traditional, statically-programmed robots) in offering care to patients beyond predicting whether not they possess conditions like CKD or breast cancer.

We further argue that biomimetic AI is needed to the extent that chronic sufferers of diseases like breast cancer and kidney disease do not want to feel helpless for most of their waking lives. Such systems would, in theory, empower patients to live on their own terms via an interface that would continuously relearn their needs, in addition to predicting and resolving crises identified via real time (i.e., “stream-processing”) processing of patient data.

4.1 Limitations

One potential concern in predicting disease presence based on patient information has to do with privacy. More specifically, any legally-sanctioned health AI or ML system would need to comply with HIPAA (Health Insurance Portability and Accountability Act of 1996) legislation. Related to this are concerns dealing with the ethics of patient information usage. Health-providing biomimetic AI’s would need to be embedded with privacy and storage settings that suit both patients and their respective, human healthcare providers.

We assert that the biomimetic AI or ML systems postulated in this paper would need to prioritize the lives of patients over all other potential concerns—including the survival of such biomimetic systems in question, if necessary. A more rigorous hierarchy of values resembling or surpassing the best human ones in existence would need to be embedded in these systems if they would be expected to face more ethically complex and demanding situations.

5 Conclusion

The potential for a “Holacratic” [6] psychoinformatic [14,15,16] organization to research solutions ([7], p. 13) to the problems engaged with in this paper remains to be concretely assessed. Holacracy–derived from the concept of holarchy (as discussed in [8, 12])—is a general governance model that “simultaneously honor[s] autonomy and enable[s] self-organization at every level within” an organization [5] (p. 39). Within a Holacratic organization, workers are “left to act…like free agents, able to shop around and accept role assignments anywhere in the organizational structure, including filling several roles in many different parts of the organization at once” (p. 39). Amidst his explication of Holacracy’s core principles, Robertson spotlighted the shoe and clothing retailer Zappos as an illustrative case example of a successful Holacratic business.

“Global solution networks” (GSNs), which have been defined by Tapscott in [6], are similar to Holacratic organizations. GSNs consist of any group of individuals with the following four features:

  • Diversity in stakeholders who collectively represent at least two of the four pillars of society (government, private sector, civil society, individuals)

  • Multinational or global presence

  • At least partial digitality with respect to its communications tools and platforms

  • Progressive goals related to the creation of “public goods” (p. 18)

GSNs have the distinct advantage of being new, thus allowing their participants to deal with problems in manners less strictly premised on partial global problem typologies such as the United Nations and The Millennium Project’s.

We contend that the Sirius Project that worked on this paper (as well as others—see, e.g., [11]) is presently a Holacratic GSN that has demonstrated potential to evolve into a “benefit corporation” (B-Corp) [13]. B-Corps are defined as new legal tools used to “create a solid foundation for long term mission alignment and value creation”. They protect said mission alignment through “capital raises and leadership changes”, which “creates more flexibility when evaluating potential sale and liquidity options, and prepares businesses to lead a mission-driven life post-IPO”. As a B-Corp, Sirius would have access to the funds needed to realize the dream of healthcare-providing biomimetic AI.