Vous êtes sur la page 1sur 9

AMERICAN JOURNAL OF PHYSICAL ANTHROPOLOGY 121:19 (2003)

Bayes Theorem in Paleopathological Diagnosis


Steven N. Byers1* and Charlotte A. Roberts2
1
Department of Anthropology, University of New Mexico, Albuquerque New Mexico 87131-1086
2
Department of Archaeology, University of Durham, Durham DH1 3LE, England, UK

KEY WORDS paleopathology; Bayes theorem; diagnosis

ABSTRACT The utility of Bayes theorem in paleo- lowing this, the sources of these prior probabilities and
pathological diagnoses is explored. Since this theorem has their accompanying problems in paleopathology are con-
been used heavily by modern clinical medicine, its useful- sidered. Finally, an application using prehistoric rib le-
ness in that eld is described rst. Next, the mechanics of sions is presented to demonstrate the utility of this
the theorem are discussed, along with methods for deriv- method to paleopathology. Am J Phys Anthropol 121:19,
ing the prior probabilities needed for its application. Fol- 2003. 2003 Wiley-Liss, Inc.

One of the most vexing problems facing paleo- placed only 42.9% of provided specimens into their
pathologists is the accurate identication of patho- correct pathological category (e.g., anomaly, trauma-
logical conditions in ancient osteological remains. As repair, or metabolic). Even worse, when identica-
discussed by numerous workers (Brothwell, 1981; tions of specic conditions were attempted (e.g.,
Ortner, 1992; Waldron, 1994; Miller et al., 1996; Pagets disease, rheumatoid arthritis), only 28.6% of
Lovell, 2000), there are many difculties associated specimens were correctly assigned. These rather dis-
with this process, because different pathological con- tressing results underscore the need for techniques
ditions have similar osteological manifestations that aid in the identication of dry bone pathologies.
(Ortner, 1992; Aufderheide and Rodriquez-Martin, In the paleopathological literature, few works de-
1998). Since the presence of pathologies has been scribe methods focused specically on differential
used to explain aspects of life in past populations diagnosis. Steinbock (1976), while devoting consid-
(Steinbock, 1976; Ortner and Putschar, 1981; Buik- erable attention to differential diagnosis, did not
stra and Ubelaker, 1994), accuracy of diagnosis is of present a general method for dealing with this prob-
paramount importance. Although many good publi- lem. Buikstra (1976) used a key diagram (also called
cations describe the visual and radiographic appear- a decision tree) as a method to aid in the identica-
ance of various diseases, traumata, and other depar- tion of tuberculosis in the spines of Caribou Eskimo.
tures from normal bone form (e.g., Brothwell and From an assemblage of defects that also could indi-
Sandison, 1967; Steinbock, 1976; Ortner and Put- cate rheumatoid arthritis, blastomycosis, brucello-
schar, 1981; Zimmerman and Kelley, 1982; Roberts sis, and other diseases, she derived the most likely
and Manchester, 1995; Aufderheide and Rodrguez- diagnosis of a pathological specimen by tracing the
Martin, 1998), very little attention is given to meth- expression of traits along different paths for differ-
ods for resolving problems in differential diagnosis. ent individuals. (Rogers et al., 1987 presented infor-
When faced with a bone containing abnormalities mation that could be used in a similar manner when
that could be caused by any number of pathological differentiating between various arthropathies.) An-
processes, it is up to the paleopathologist to use his other method, typied by Blackman et al. (1991),
or her best judgment to correctly identify the caus- involves making a decision table to relate diseases to
ative agent. their concomitant pathological signs and other rele-
The lack of methods for untangling conicting in- vant data (e.g., demographics). This is constructed
formation has often resulted in fairly strong dis- by listing signs of pathological conditions (or other
agreements among researchers. Miller et al. (1996) data that can be used in a diagnosis) down the
showed that even persons experienced in the prac-
tical aspects of skeletal biology, paleopathology, and
medicine often offer incorrect judgments, and/or *Correspondence to: Steven N. Byers, Department of Anthropology,
may be unsure of a precise pathological classica- University of New Mexico, Albuquerque, NM 87131.
tion. Using dry bone lesions of known etiology, these
Received 24 August 2001; accepted 30 May 2002.
authors compiled data on the accuracy of diagnoses
made by persons attending workshops in 1990 and DOI 10.1002/ajpa.10164
1991 at the annual meetings of the Paleopathology
Association. Their results showed that attendees

2003 WILEY-LISS, INC.


2 S.N. BYERS AND C.A. ROBERTS

left-hand side of a table, while the different patho- TABLE 1. Pathological conditions explored using Bayes
logical entities that contain these signs are arranged theorem or Bayesian logic
across the top as column headings. The expression of Pathological condition Reference
each sign or other data for a condition is entered in Alzheimers disease Prince, 1996
the appropriate conuence of column and row, and a Ankylosing spondylitis Diffey et al., 1985
diagnosis is determined by the column that best Appendicitis Edwards and Davies, 1984
describes the skeleton(s) under study. Unfortu- Asthma Perpina et al., 1993
Bone tumors Lodwick et al., 1963
nately, these are the only examples from the litera- Cancer Du Boulay et al., 1977; Kahn et
ture involving methods of paleopathological diagno- al., 1997; Lind and Singer,
sis. Thus, this area is in need of concentrated 1986; Montironi et al., 1994,
attention if a greater battery of procedures is to be 1998; Pastor et al., 1997
Colonic diseases Siles et al., 1997; van de
available to the osteological researcher. One source Merwe, 1983
of such methods is the techniques used by clinicians Coronary artery disease Diamond et al., 1980
to diagnose pathologies in living persons. Cushings Syndrome Nugent et al., 1964
Gallbladder disease Haddawy et al., 1994
In the medical literature, most solutions to the prob- Hearing loss Phaneuf et al., 1985
lem of diagnosis can be subsumed under the overall Heart disease Warner et al., 1961, 1964;
heading of decision support systems (DSSs). In a good Reale et al., 1968
overview, Degoulet and Fieschi (1997) distinguished Hypertension Blinowska et al., 1991;
Elijovich and Laffer, 1992
four methodological approaches underlying these sys- Jaundice Malchow-Moller et al., 1986
tems: mathematics, expert systems, neural networks, Liver and biliary Begon et al., 1979
and probability. Mathematical models involve calcu- diseases
lating functional relationships between physical Lupus Somogyi et al., 1993
Mental retardation Mani et al., 1997
and/or physiological parameters of interest (e.g., blood Multiple sclerosis Blinowska et al., 1992
vessel size and ow rate). Because these functions do Osteomyelitis Worbel and Connolly, 1998
not calculate categorical variables that would repre- Pneumonia Aronsky and Haug, 1998
Rheumatological Bernelot Moens and van der
sent pathological conditions, their applicability to di- disorders Korst, 1991, 1992
agnosis is limited. Expert systems are computer appli- Rheumatoid arthritis Bernelot Moens et al., 1992
cations that use articial intelligence languages to Streptococcal infections Poses et al., 1986
mimic the methods used by the most skillful diagnos- Thyroid disease Overall and Williams, 1963
Viral infections Berg, 1981
ticians, while neural networks are computer systems
that mimic the structure and function of the human
brain. The effectiveness of these two approaches is still
being researched. Finally, probabilistic methods use found that the theorem was correct 9193% of the
various observed frequencies of pathological condi- time, which is a signicant increase over the 50%
tions in conjunction with their signs and/or symptoms probability due to chance alone. In addition, where
to develop probabilities of these conditions in patients reported, the theorem performs as well as, or better
presenting similar clinical pictures. A number of sta- than, the initial diagnosis by physicians. Since the
tistical procedures are available in this method, in- purpose of both modern clinical medicine and paleo-
cluding Bayes theorem, logistic regression and/or dis- pathology is accurate diagnosis, the information in
crimination, and linear discriminant analysis. Tables 1 and 2 indicates that this method should be
Of the four methodological approaches, probabil- useful in paleopathology. Therefore, this paper ex-
ity models using Bayesian theory and especially plores how Bayes theorem could be applied to pa-
Bayes theorem are the most widely employed. Table leopathology. The issues and problems surrounding
1 presents a partial list drawn from the medical its application will be discussed, and an example
literature of pathological conditions to which clini- given. It is hoped that the presentation here will
cians have applied Bayes theorem or Bayesian the- stimulate further research into this promising
ory to the problem of diagnosis. As can be seen, even method.
an incomplete list is extensive, indicating that the
BAYES THEOREM
theory has been widely studied by the medical es-
tablishment. Furthermore, Bayes theorem has been Bayes theorem describes how knowledge of prior
shown to be as good as, or better than, both clinical probabilities can be used to calculate the probability
diagnoses and guessing. Table 2 presents results of unknown events in many subject areas, including
from a sample of medical studies where Bayes the- the sciences, the humanities, and even games of
orem was used to diagnose a number of conditions chance. In problems involving the diagnosis of mod-
whose correct identication was obtained from ern disease, the prior probabilities are both the
complete testing. As can be seen, in almost all stud- prevalences of pathological conditions and the like-
ies, Bayes theorem predicts the correct diagnoses lihoods of signs found within those conditions.
better (usually much better) than expected from Prevalences are the frequencies with which patho-
chance alone. For example, when applied to the logical conditions (i.e., infectious diseases, nutri-
problem of differentiating appendicitis from other tional deciencies, trauma, and any other departure
abdominal problems, Edwards and Davies (1984) from normal structure and function) occur in popu-
DIAGNOSIS USING BAYES THEOREM 3
TABLE 2. Accuracy of Bayes theorem in selected clinical studies1

General pathological Number of % correct diagnosis


condition specic conditions Clinicians Bayes theorem Reference
Congenital heart disease 33 73.2 81.4 Warner et al., 1961
Congenital heart disease 94 73.2 81.8 Reale et al., 1968
Polycythemic states 5 6576 95 Bishop and Warner, 1969
Intrathoracic radiographs 15 nr 53 Alperovitch and Lellouch,
1974
Thoracic conditions 3 nr 73 McNeil and Sherman,
1978
Gastrointestinal bleeding 2 nr 4969 Ohmann et al., 1988
Acute abdominal pain 9 76 74 Gammerman and
Thatcher, 1991
Liver disease 40 nr 6477 Croft and Mochol, 1974
Appendicitis 2 7689 9193 Edwards and Davies,
1984
1
nr, not reported.

lations. Likelihoods are the frequencies with which ber of signs used has varied from fewer than 20 (e.g.,
signs (i.e., visually observable characteristics, symp- Sherman, 1978; Bishop and Warner, 1969) to
tomatic complaints, the results of medical tests, lo- around 30 (e.g., Alperovitch and Lellouch, 1974; Fry-
cation of abnormalities within the body, and any back, 1978; Gammerman and Thatcher, 1991) and
other attribute that can be used to identify a patho- as high as 50 (e.g., Warner et al., 1961).
logical condition) appear within the conditions for Once these values are determined, Bayes theorem
which the prevalence is known. can aid in diagnosis in the following manner. First,
From these statistics, the simplest form of Bayes those signs found in a patient are identied, and
theorem can be used to calculate the probability of their prior likelihoods are entered into equation (1).
the presence or absence of an ailment, given its This equation is then calculated for each pathologi-
prevalence and the likelihood of a single sign occur- cal condition for which prevalences exist. The condi-
ring in people suffering from that condition. In its tion with the highest resulting probability would be
general form, it can be extended for multiple dis- the most likely diagnosis. Given this, it seems rea-
eases and multiple signs so that the probability of a sonable that Bayes theorem has applications to di-
pathological condition can be calculated as: agnosing osteopathological conditions. Since modern
medicine has demonstrated its efcacy and since
PC i S 1 , S 2 . . . S s
both paleopathologists and clinicians have similar
PC i LS 1 C i S 2 C i . . . S S C i goals, the theorem should help meet the aforemen-
(1) tioned need for methods in paleopathological diag-
PC LS C S C
C
nosis. All that needs to be done is to derive the
i 1 i 2 i . . . S S C i
necessary prior probabilities, and equation (1) could
i1
be applied.
where: Prevalences for each pathological condition can be
calculated according to the following formula:
P(Ci|S1, S2 . . . Ss) Probability of pathologi-
cal condition i, given N Ci
the presence of signs 1, PC i (2)
N
2, etc.
P(Ci) Prevalence of pathologi- where:
cal condition i
L(S1|Ci, S2|Ci . . . P(Ci) Prevalence of pathological condition i
ScCi) Likelihood of sign 1, NCi Number of people with pathological condi-
sign 2, etc. in patholog- tion i
ical condition i N Total number of people in the population
From equation (1), it can be seen that any number of As can be seen, the number of individuals suffer-
conditions (Cc) or signs (Ss) can be used in the cal- ing from a condition is simply divided by the total
culation of the probability of a disorder. In the med- number of individuals from the population on which
ical literature, the number of conditions analyzed the statistic is based. This calculation is carried out
has varied from 1 (e.g., Ohmann et al., 1988) to for each pathological condition.
almost 100 (e.g., Reale et al., 1968), with most re- The computation of likelihoods follows a similar
searchers limiting their study to under 30 (e.g., Al- course. This statistic is computed by dividing the
perovitch and Lellouch, 1974; McNeil and Sherman, number of individuals having both a sign and a
1978; Gammerman and Thatcher, 1991). The num- pathological condition by the total number of indi-
4 S.N. BYERS AND C.A. ROBERTS

viduals suffering from that pathological condition. ple of the midwestern United States in the earlier
Thus: part of the 1900s. Newer skeletal collections have
comparable problems with sample bias, i.e., they are
NS j C i
LS j C i (3) derived from specic regions in the United States.
N Ci Similarly, although modern health statistics from
where: Third World countries may be more representative
of past populations because they derive from people
L(Sj|Ci) Likelihood of sign j in condition i living a traditional lifestyle without access to mod-
N(Sj|Ci) Number of persons exhibiting sign j in ern therapies, the impact of diseases when people
condition i were rst exposed to them may have been different
NCi Number of persons with condition i than they are today (Brothwell, 1967; Ortner and
Pustchar, 1981). This would bias prevalences and
This calculation is carried out on all signs for a likelihoods calculated from these statistics. In sum,
pathological condition. all sources of prior probabilities have problems with
As can be seen, to apply Bayes theorem to paleo- representativeness; however, this simply character-
pathology requires only that the prevalences and izes the state of the science today. For example, the
likelihoods be available for input to equation (1). Terry and Todd Collections provide much of the
However, a number of issues and problems sur- information used by human skeletal biologists in
rounding their derivation require explanation. osteological studies; this makes them the source of
PREVALENCES AND LIKELIHOODS IN standards in physical anthropology, whether or not
PALEOPATHOLOGY they are representative of the population at large.
Also, as will be discussed below, the effect of this
In modern medicine, most researchers calculate problem is not so critical as to invalidate the
prevalences and likelihoods from patients seen in a method.
clinical setting for whom positive diagnoses have The next issue to consider is the accuracy of the
been obtained (e.g., Ledley and Lusted, 1959; diagnoses in these sources. As pointed out by Roth-
Warner et al., 1961; Reale et al., 1968; Bishop and schild et al. (1990) when studying the Terry Collec-
Warner, 1969; Croft and Machol, 1974; Alperovitch tion, modern criteria for the identication of patho-
and Lellouch, 1974; Norusis and Jacquez, 1975; logical conditions were unavailable in earlier times.
Starmer and Lee, 1976; Fryback, 1978; McNeil and Therefore, prevalences and likelihoods derived from
Sherman, 1978; Ohmann et al., 1988; Gammerman early health statistics in industrialized countries, or
and Thatcher, 1991). Since similar information is calculated from older skeletal series, can be ques-
not available to paleopathologists, these prior prob- tioned. For example, although some support the di-
abilities have to be derived elsewhere. One obvious agnoses seen in the records for the Terry Collection
source is health statistics on living populations, both (David Hunt, personal communication), others have
old data collected by industrialized countries from noted that a number of these are laughable (Stan
the 1800s and early 1900s as well as new statistics Rhine, personal communication). Despite these pos-
published by modern health agencies. Another sible problems, it seems unlikely that all (or even
source is data that could be gathered from the col- most) of the diagnoses in early health statistics or
lections of human skeletons, such as the Terry and skeletal collections are incorrect; thus, their use in
Todd Collections and the newer but smaller collec- computing prior probabilities is supported. Another
tions at places such as the University of New Mexico problem with diagnoses is that health statistics and
and the University of Tennessee, Knoxville. skeletal collection records list the presence of a dis-
Before any of these are used to compute prior ease usually only if it was the cause of death. Dis-
probabilities, a number of issues must be consid- eases suffered during life but not causing death are
ered. The rst is, how representative of prehistoric noted only infrequently (e.g., Terry Collection), if at
populations are the health statistics and data that all (e.g., modern health statistics).
are derived from skeletal collections? It seems rea- From the above discussion, it is evident that ob-
sonable that the older the data (e.g., old health sta- taining prevalences and likelihoods that are appli-
tistics, data from the Terry and Todd Collections), cable to past populations is a difcult endeavor, at
the more likely they are to reect the prehistoric best. Although the problems seem insurmountable,
condition because the effects of modern therapies there is evidence that their resolution may not need
would not mitigate (or eliminate) the defects used in to meet all of the criteria demanded by science. In a
the determination of prevalences and likelihoods. potentially devastating article, Miettinen and Caro
Thus, the people on whom these prior probabilities (1994) point out unsolvable epistemological prob-
would be computed are more analogous to past pop- lems with the calculation of these prior probabilities
ulations than the people represented in newer in modern clinical medicine. Their thesis is that
sources. However, these samples contain biases of the application of Bayes theorem to the identica-
unknown effect, e.g., old health statistics are avail- tion of disease is based on untenable premises. At
able only from industrialized countries, and the the same time, however, they openly admit that
Terry and Todd Collections were derived from peo- feasible diagnostic probabilities are calculable. This
DIAGNOSIS USING BAYES THEOREM 5
is encouraging for paleopathologists. If modern cli- ence the rate at which persons will suffer patholog-
nicians can compute useful probabilities (as seen by ical conditions. For example, the frequency of genet-
the number of pathological conditions in Table 2) ically induced diseases such as sickle-cell anemia
despite underlying problems with prevalences and and thalassemia, which can have osteological man-
likelihoods, Bayes theorem almost certainly has ap- ifestations (Cooley et al., 1927; Sebes and Diggs,
plication to the identication of osteological mani- 1979; Whipple and Bradford, 1932), would cause
festations of diseases in past populations. In addi- prevalences in some areas of the world to be differ-
tion, it has been shown that the theorem is robust, ent from those in other areas. Similarly, environ-
since it yields useful results even when the assump- mental factors such as climate and weather, which
tions on which it is based (e.g., independence of long have been recognized to be correlated with the
signs; see below) are violated in a limited manner. appearance of diseases and often are seen to be
Thus, despite the difculties surrounding prior related to seasonality (Yan, 2000; De Garine, 1993;
probabilities, Bayes theorem deserves the attention Lukacs and Walimbe, 1998; Steinbock, 1976; Ortner
of paleopathologists. and Putschar, 1981; Aufderheide and Rodrguez-
No matter which source is chosen for prevalences Martin, 1998), also make the calculation of this sta-
and likelihoods, a number of issues surrounding tistic dependent on geographic location. Even cul-
these statistics must be considered. For prevalences, tural factors such as dietary elements and their
four of these are of major importance. First, the list combinations (e.g., Larsen, 1995; El-Najar et al.,
of pathological conditions for which prevalences are 1975), or simply the aggregation of people in towns
calculated must be exhaustive, i.e., all possible and cities (Brothwell, 1967; Howe, 1997; Schell and
prevalences must be available for use in Bayes the- Ulijaszek, 1999), would affect the frequency of dis-
orem. In the clinical literature, this problem is eases. Finally, many workers (e.g., Steinbock, 1976;
solved by identifying the diseases specically being Ortner and Putschar, 1981; Roberts and Manches-
studied and then creating a normal or other cat- ter, 1995; Aufderheide and Rodrguez-Martin, 1998;
egory for all other conditions (e.g., Warner et al., Grauer and Stuart-Macadam, 1998; Pollard and
1961; McNeil and Sherman, 1978). Another method Hyatt, 1999) indicate that demographic factors such
for dealing with this is to restrict the number of as sex and age are correlated with pathological con-
diseases being examined (e.g., Alperovitch and Lel-
ditions, which then would affect the calculation of
louch, 1974; Gammerman and Thatcher, 1991), so
prevalences.
that diagnoses are attempted only on patients suf-
Likelihoods have three issues to consider. The
fering from specic pathologies.
rst is the simple identication of signs. Ortner
A second issue is that prevalences are calculated for
(1994) delineated two factors that are important
conditions that are mutually exclusive. Thus, a person
cannot have more than one of the C conditions ap- when describing a sign of a pathological condition:
pearing in equation (1). Unfortunately, despite the fact what is its nature, and where is it located? To these
that patients suffering from multiple diseases are part can be added demographic characteristics such as
of the clinical world, most workers simply ignore this ancestral group, age, and sex (Ortner and Pustchar,
problem (e.g., Warner et al., 1961; Reale et al., 1968; 1981). Unfortunately, present methods for describ-
Bishop and Warner, 1969; Alperovitch and Lellouch, ing the nature of abnormalities are not yet as well-
1974; Gammerman and Thatcher, 1991), while only a developed as paleopathologists might wish (Ortner,
few (e.g., Reale et al., 1968) create disease categories 1991, 1992, 1994). Generally four characteristics,
which are combinations of individual conditions. and their combinations, have been presented as in-
A third issue involves the sum of the prevalences. dicators of a pathologically induced change (Ortner
Traditionally, the prevalences of all diseases add up and Putschar, 1981; Ortner, 1992, 1994): abnormal
to 1 because the P(Ci)s are computed from a popu- bone loss, abnormal bone gain, abnormal shape, and
lation of individuals suffering from any number of abnormal size. In addition to these, the amount of
pathological conditions of interest. However, despite bone that is affected, either in absolute measure or
the widespread use of this standard, it is not neces- percent of the total bone, is similarly important
sary for the proper application of Bayes theorem. (Buikstra and Ubelaker, 1994).
Since this formula describes the relationship be- Fortunately, the other feature of description (i.e.,
tween probabilities, any values of prevalence that location of lesions) is fairly easy to dene. This in-
are considered valid can be used. The only require- volves naming the bone(s) manifesting the condi-
ment is that the relationship between the frequency tion(s), and the placement of defects within it/them
of diseases must be maintained (i.e., if one disease is (Ortner, 1992). Buikstra and Ubelaker (1994) thor-
twice as common as another, the prevalence of the oughly described this latter factor in a typical long
rst must be twice that of the second). Similarly, if bone as involving side (i.e., right, left), section (i.e.,
only the rank order of prevalences is known, Bayes diaphysis, metaphysis, epiphysis), and aspect (i.e.,
theorem still can be applied (for an example, see medial vs. lateral, proximal vs. distal, anterior vs.
Horbar, 1983). posterior, and circumferential). Similar descriptions
The nal complication affecting prevalences are are available for other bones of the skeleton, and
intrinsic and extrinsic population factors that inu- more detail could be useful.
6 S.N. BYERS AND C.A. ROBERTS

The second issue affecting likelihoods is the as-


sumption of independence, i.e., the occurrence of one
sign does not depend on the presence of another.
Since many co-occur, unusually high or low proba-
bilities may result from Bayes theorem calculations,
thereby causing misclassication of diseases. In the
medical literature, this problem has been ap-
proached either by ignoring dependencies (called
simple Bayes by Gammerman and Thatcher, 1991)
or by accounting for these in the calculation of dis-
ease probabilities (termed proper Bayes by Gam- Fig. 1. Osteoproliferative lesion on internal surface of a rib
merman and Thatcher, 1991). Some studies show from prehistoric Louisiana.
that misclassication goes down with proper
Bayes (Norusis and Jacquez, 1975; Russek et al.,
1983; Ohmann et al., 1988; Chard, 1989), while oth- and fragmentation of osteological material had oc-
ers show an increase in misclassication when de- curred. Thus, there were no associated skeletal ele-
pendencies are taken into account (Fryback, 1978; ments to aid in the diagnosis or any opportunity to
Gammerman and Thatcher, 1991). On the whole, look at the distribution pattern of lesions.
however, the data indicate that misclassication Roberts et al. (1994) researched the incidence of
rates between simple and proper Bayes are not similar osteoproliferative lesions in some 380 skele-
great when a small number of signs (e.g., fewer than tons from the Terry Collection. Their study revealed
10) is used (see Norusis and Jacquez, 1975; Fryback, lesion frequencies in persons whose deaths were
1978; Russek et al., 1983; Ohmann et al., 1988; recorded as being due either to tuberculosis (Tub),
Chard, 1989). Thus, although the assumption of in- other pulmonary disease (Oth Pul), or nonpulmo-
dependence between signs is theoretically incorrect, nary disease (NonPul). These conditions met the
Bayes theorem is robust enough to be of value in criteria discussed above for the calculation of preva-
diagnosis, even when this assumption is violated in lences, i.e., they represented a mutually exclusive
a limited manner. and exhaustive list. In addition, their frequencies
The last problem involves determining which added up to 1.0, and the disease of interest was
signs to employ in the calculation of Bayes theorem. identied (in this case, tuberculosis), while diseases
In modern clinical medicine where many signs are of lesser interest were placed into other categories.
considered, some researchers check rst for causally Also, the precept by Ortner (1994, p. 6) that the
related (dependent) signs and use only the one from nature of the abnormality must rst be specied
the group with the highest likelihood (e.g., Warner was satised by the description osteoproliferative
et al., 1961; Ohmann et al., 1988; Gammerman and lesions. Likelihoods then would be based on fre-
Thatcher, 1991). Others only use signs that occur quencies concerning location as well as incidence by
more often in one disease than another (e.g., McNeil age, sex, and other demographic characteristics. In
and Sherman, 1978; Norusis and Jacquez, 1975). this case, Roberts et al. (1994) indicated that age,
This information indicates that only the likelihoods side, location in thorax, and affected surface of ribs
from uncorrelated (independent) signs that have fre- were useful for distinguishing the three cause-of-
quencies that are either high (i.e., approaching 1) or death categories. Also, the independence between
low (i.e., approaching 0) should be used in the cal- signs was not important, since there were fewer
culation of the theorem, while those whose frequen- than 10, making the search for correlation unneces-
cies are correlated and nearly equal in all diseases sary.
should be avoided. However, as discussed above, this When calculating Bayes theorem, the usual
issue can be ignored if fewer than 10 signs are being method is to place the prior probabilities into a table
considered. with diseases arranged as rows, and signs as col-
PALEOPATHOLOGICAL EXAMPLE umns (Warner et al., 1961; Vishnevskiy et al., 1973).
In this manner, the likelihoods of signs within dis-
As an example of how Bayes theorem could be eases are located at the intersection of the disease
applied to paleopathology, one of the authors row with the sign column. Table 3 presents a sum-
(S.N.B.) discovered rib fragments with osteoprolif- mary of the data of Roberts et al. (1994) for the three
erative lesions (Fig. 1) in a collection of bone exca- causes of death. These statistics show that there are
vated by Czajkowski (1934) from Little Woods, a no characteristics that occur only in one disease (i.e.,
Native American site in southern Louisiana dated likelihood 1.0) and not in the others. Although this
500 250 BC. Since tuberculosis has been hypothe- would allow for a more condent diagnosis, this sit-
sized to occur prehistorically in the Americas (for a uation is so rare that it should not be expected.
list of sources, see Powell, 1992), albeit later in time, Second, age at death is a useful characteristic for
it was reasonable to question if these lesions repre- determining disease (e.g., persons between ages
sented that disease. Unfortunately, the ribs were 26 35 are six times more likely to die of tuberculosis
from an ossuary where considerable disarticulation than from a nonpulmonary disease). Although it is
DIAGNOSIS USING BAYES THEOREM 7
TABLE 3. Prior probabilities for application of bayes theorem to osteoproliferative lesions
Position of lesions within
Age at death Side of lesion cage Position of lesions on ribs
Cause of death Prevalence 1525 2635 3645 4655 56 One Both Low Mid High All Anterior Lateral Posterior All

Tuberculosis 0.421 0.155 0.316 0.226 0.142 0.161 0.548 0.452 0.019 0.382 0.076 0.204 0.076 0.064 0.057 0.611
Other pulmonary 0.137 0.100 0.100 0.160 0.220 0.420 0.760 0.240 0.020 0.529 0.039 0.137 0.200 0.040 0.080 0.440
Nonpulmonary 0.442 0.049 0.049 0.221 0.221 0.436 0.521 0.479 0.069 0.500 0.050 0.163 0.282 0.043 0.049 0.399

theoretically incorrect to include such frequencies ability of nonpulmonary cause of death is only P
(see Miettenen and Caro, 1994), they can be used in 0.679). However, in other cases, stronger probabilities
calculations because of the robust nature of the the- result; for example, if these same lesions were found
orem discussed above. along the entire surface of both sides of the upper ribs
Since the age, side, and position within the rib of a skeleton from an individual who was around 30
cage cannot be determined for the individual(s) from years old at the time of death, the probabilities from
Little Woods, the only usable likelihood is the af- Bayes theorem given these four signs would be:
fected rib surface. Because these all are located an- P(Tub|4 Signs) 0.914, P(Oth Pul|4 Signs) 0.018,
teriorally (Ant), equation (1) for the three causes of and P(NonPul|4 Signs) 0.068, resulting in a better
death are: case for tuberculosis being the cause of the lesions.
Last, given the paucity of data used in the above cal-
PTubAnt
culations, diagnosis based solely on the computed
.421*.076 probabilities would be unwarranted. This situation
would be true even in the hypothetical example where
.421*.076 .137*.200 .442*.282
the probabilities were more compelling. It should
.032 .032 never be construed that the results of Bayes theorem
.174 (4)
.032 .027 .125 .184 would be the only basis for developing a diagnosis;
rather, all data should be considered before attempt-
POth PulAnt
ing to identify a pathological condition.
.137*.200 As a way of reinforcing this last point, consider the
following analysis. Since the question being ex-
.421*.076 .137*.200 .442*.282
plored here is whether or not the osteoproliferative
.027 .027 lesions are due to tuberculosis, other signs of this
.147 (5)
.032 .027 .125 .184 disease should be evident in the skeletal material
despite their disarticulation. According to Ortner
PNonPulAnt and Putschar (1981), the primary osteological man-
.442*.282 ifestation of tuberculosis is cavitating lesions in the
hemopoietic marrow, with little reactive bone forma-
.421*.076 .137*.200 .442*.282
tion. This defect is found most often in the bodies of
.125 .125 vertebrae, especially the lumbar region, where it
.679 (6)
.032 .027 .125 .184 usually affects two or more bones in the same ver-
tebral column. In the joints, the hip and knee are
These calculations reveal several important points. affected most often, with destruction of the articular
First, there is a higher probability that the lesions are surfaces followed by fusion (ankylosis). Among 95
associated with a nonpulmonary cause of death than complete or nearly complete vertebral bodies from
with tuberculosis. This is surprising, because Roberts Little Woods, there is only one which exhibits a lytic
et al. (1994) showed that tuberculosis is a highly prob- lesion that approximates the cavitations expected in
able cause for these osteoproliferations. However, tuberculosis, and none of the joints show destruction
since they are not pathognomonic for this infection and ankylosis. Thus, a diagnosis of tuberculosis for
(other pulmonary diseases such as pneumonia and these rib fragments is unwarranted, and the pres-
neoplastic disease also exhibit these lesions), the re- ence of this disease at the site cannot be proven with
sult here is reasonable. Second, these calculations dou- these data. Unfortunately, because nonpulmonary
ble the probability of correct diagnosis over that due to causes of death were not investigated further by
chance. That is, all things being equal, there is only a Roberts et al. (1994), nothing additional can be as-
0.333 probability that a nonpulmonary disease caused certained as to the cause of these lesions.
the lesions. Third, the calculations are predicated on
the applicability of Terry Collection frequencies to the
CONCLUSIONS
prehistoric peoples of southern Louisiana. The as-
sumption of representativeness of sample prior prob- To summarize, Bayes theorem has been used suc-
abilities has already been discussed; however, for the cessfully by the medical profession to aid in the
purpose of illustration, the sample bias of the Terry diagnosis of disease in a clinical setting. Although
Collection is ignored. Fourth, sometimes probabilities physicians have the advantage of a larger range of
are not as persuasive as one would like (e.g., the prob- tests as well as therapies whose effectiveness can be
8 S.N. BYERS AND C.A. ROBERTS

used to support a diagnosis, the method is still ap- Blinowska A, Verroust J, Malapert D. 1992. Bayesian statistics
plicable to paleopathology. Despite the fact that as applied to multiple sclerosis diagnosis by evoked potentials.
Electromyogr Clin Neurophysiol 32:1725.
there are difculties in obtaining reasonable preva- Brothwell DR. 1967. The bio-cultural background to disease. In:
lences and likelihoods, the derivation of these sta- Brothwell DR, Sandision AT, editors. Diseases in antiquity.
tistics would make it possible to calculate the prob- Springeld, IL: C.C. Thomas. p 56 68.
ability of any number of pathological conditions for Brothwell DR. 1981. Digging up bones, 3rd ed. Ithaca, NY: Cor-
nell University Press.
skeletons manifesting osteological abnormalities.
Brothwell DR, Sandison AT. 1967. Diseases in antiquity. Spring-
Prevalence and likelihood statistics could be derived eld, IL: C.C. Thomas.
by a comprehensive study of all pathologies in the Buikstra JE. 1976. The caribou Eskimo: general and specic
Terry and Todd Collections as well as other smaller disease. Am J Phys Anthropol 45:351368.
skeletal series, using a data-gathering instrument Buikstra JE, Ubelaker DL. 1994. Standards for data collection
from human skeletal remains. Arkansas Archeological Survey
that includes, and goes beyond, those characteristics research series no. 44. Fayetteville, AR: Arkansas Archeologi-
described by Buikstra and Ubelaker (1994), Lovell cal Survey.
(2000), or Byers (2002). Additionally, information on Byers SN. 2002. A model for the diagnostic process in paleopa-
prevalences could be obtained from modern health thology. Paleopathol Newslett 117:1118.
statistics, especially from the Third World, as a Chard T. 1989. The effect of dependence on the performance of
Bayes theorem: an evaluation using computer simulation.
means of checking the applicability of skeletal series Comput Methods Programs Biomed 29:1519.
data to the world at large. Information gathered in Cooley TB, Witner ER, Lee P. 1927. Anemia in children. Am J Dis
this manner could be used to develop a computerized Child 34:347363.
diagnosis system (S.N.B. has a prototype of such a Croft DJ, Machol RE. 1974. Mathematical methods in medical
system, which is available by e-mailing him at diagnosis. Ann Biomed Eng 2:69 89.
Czajkowski JR. 1934. Preliminary report of archeological excava-
j708@unm.edu), which would calculate probabilities tions in Orleans Parish. Louisiana Conservation Rev 4:1218.
that could support or refute diagnoses developed De Garine I. 1993. Culture, seasons and stress in two traditional
using more traditional methods. This would provide African cultures (Massa and Mussey). In: Ulijaszek SJ, Strick-
more information for the diagnostic process, thereby land SS, editors. Seasonality and human ecology. Cambridge:
University Press. p 184 201.
increasing the accuracy with which paleopathologi- Degoulet P, Fieschi M. 1997. Introduction to clinical informatics.
cal conditions are identied. New York: Springer.
Diamond GA, Forrester JS, Hirsch M, Staniloff HM, Vas R,
LITERATURE CITED Berman DS, Swan HJ. 1980. Application of conditional proba-
Alperovitch A, Lellouch J. 1974. Methods for aiding medical de- bility analysis to the clinical diagnosis of coronary artery dis-
cision: application to diagnosis of round intra-thoracic x-ray ease. J Clin Invest 65:1210 1221.
picture. Comput Biomed Res 7:127141. Diffey BL, Pal B, Gibson CJ, Clayton CB, Grifths ID. 1985.
Aronsky D, Haug PJ. 1998. Diagnosing community-acquired Application of Bayes theorem to the diagnosis of ankylosing
pneumonia with a Bayesian network. Proc AMAI Symp 1998: spondylitis from radioisotope bone scans. Ann Rheum Dis 44:
632 636. 667 670.
Aufderheide AC, Rodrguez-Martin C. 1998. The Cambridge en- Du Boulay GH, Teather D, Harling D, Clarke G. 1977. Improve-
cyclopedia of human paleopathology. Cambridge: Cambridge ment in the computer-assisted diagnosis of cerebral tumours.
University Press. Br J Radiol 50:849 854.
Begon F, Lockhart AM, Metreau JM, Dhumeaux DA. 1979. Com- Edwards FH, Davies RS. 1984. Use of a Bayesian algorithm in the
puter-aided system for the diagnosis of hepato-biliary diseases. computer-assisted diagnosis of appendicitis. Surg Gynecol Ob-
A comparison with the performance of physicians. Med Inf stet 158:219 222.
(Lond) 4:35 42. Elijovich F, Laffer CL. 1992. Bayesian analysis supports use of
Berg AO. 1981. Case-control diagnosis and Bayesian inference in ambulatory blood pressure monitors for screening. Hyperten-
common viral infections. J Fam Pract 12:10171021. sion [Suppl] 19:268 272.
Bernelot Moens HJ, van der Korst JK. 1991. Comparison of El-Najar MY, Lozoff B, Ryan DJ. 1975. The paleoepidemiology of
rheumatological diagnoses by a Bayesian program and by phy- porotic hyperostosis in the American Southwest: radiological
sicians. Methods Inf Med 30:187193. and ecological considerations. AJR Radium Ther Nucl Med
Bernelot Moens HJ, van der Korst JK. 1992. Development and 125:918 924.
validation of a computer program using Bayes theorem to Fryback DG. 1978. Bayes theorem and conditional nonindepen-
support diagnosis of rheumatic disorders. Ann Rheum Dis 51: dence of data in medical diagnosis. Comput Biomed Res 11:
266 271. 423 434.
Bernelot Moens HJ, Hirshberg AJ, Claessens AA. 1992. Data- Gammerman A, Thatcher AR. 1991. Bayesian diagnostic proba-
source effects on the sensitivities and specicities of clinical bilities without assuming independence of symptoms. Methods
features in the diagnosis of rheumatoid arthritis: the relevance Inf Med 30:1522.
of multiple sources of knowledge for a decision-support system. Grauer AL, Stuart-Macadam P. 1998. Sex and gender in paleo-
Med Decis Making 12:250 258. pathological perspective. Cambridge: Cambridge University
Bishop CR, Warner HR. 1969. A mathematical approach to med- Press.
ical diagnosis: application to polycythemic states utilizing clin- Haddawy P, Kahn CE Jr, Butarbutar M. 1994. A Bayesian network
ical ndings with values continuously distributed. Comput model for radiological diagnosis and procedure selection: work-up
Biomed Res 2:486 493. of suspected gallbladder disease. Med Phys 21:11851192.
Blackman J, Allison MJ, Aufderheide, Oldroyd N, Steinbock RT. Horbar JD. 1983. Revising ranked probabilities: a Bayesian ap-
1991. Secondary hyperparathyroidism in an Andean mummy. proach to incomplete knowledge. Comput Biomed Res 16:367
In: Ortner DJ, Aufderheide AC, editors. Human paleopathol- 377.
ogy: current syntheses and future options. Washington, DC: Howe GM. 1997. People, environment, disease and death. Cardiff:
Smithsonian Institution Press. p 291296. University of Wales Press.
Blinowska A, Chatellier G, Bernier J, Lavril M. 1991. Bayesian Kahn CE Jr, Roberts LM, Wang K, Jenks D, Haddawy P. 1997.
statistics as applied to hypertension diagnosis. IEEE Trans Construction of a Bayesian network for mammographic diag-
Biomed Eng 38:699 706. nosis of breast cancer. Comput Biol Med 27:19 29.
DIAGNOSIS USING BAYES THEOREM 9
Larsen CS. 1995. Biological changes in human populations with Phaneuf R, Hetu R, Hanley JA. 1985. A Bayesian approach for
agriculture. Annu Rev Anthropol 24:185213. predicting judged hearing disability. Am J Ind Med 7:343352.
Ledley RS, Lusted LB. 1959. Reasoning functions of medical Pollard T, Hyatt SB. 1999. Sex, gender and health. Cambridge:
diagnosis. Science 130:9 21. Cambridge University Press. Poses RM, Cebul RD, Collins M,
Lind SE, Singer DE. 1986. Diagnosing liver metastases: a Bayes- Fager SS. 1986. The importance of disease prevalence in trans-
ian analysis. J Clin Oncol 4:379 388. porting clinical prediction rules. The case of streptococcal phar-
Lodwick GS, Haun DL, Smith WE, Keller RF, Roberston ED. yngitis. Ann Intern Med 105:586 591.
1963. Computer diagnosis of primary bone tumors: a prelimi- Powell ML. 1992. Health and disease in the late prehistoric
nary report. Radiology 80:273. southeast. In: Verano JW, Ubelaker DH, editors. Disease and
Lovell NC. 2000. Paleopathological description and diagnosis. In: demography in the Americas. Washington, DC: Smithsonian
Katzenbery MA, Saunders SR, editors. Biological anthropology Institution Press. p 4153.
of the human skeleton. New York: Wiley-Liss. p 217248. Prince MJ. 1996. Predicting the onset of Alzheimers disease
Lukacs JR, Walimbe SR. 1998. Physiological stress in prehistoric using Bayes theorem. Am J Epidemiol 143:301308.
India: new data on localized hypoplasia of primary canines Reale A, Maccacaro GA, Rocca E, DIntino S, Gioffre PA, Vestri A,
linked to climate and subsistence. J Archaeol Sci 25:571585. Motolese M. 1968. Computer diagnosis of congenital heart dis-
Malchow-Moller A, Thomsen C, Matzen P, Mindeholm L, Bjerre- ease. Comput Biomed Res 1:533549.
gaard B, Bryant S, Hilden J, Holst-Christensen J, Johansen Roberts C, Manchester K. 1995. The archaeology of disease.
TS, Juhl E. 1986. Computer diagnosis in jaundice. Bayes rule Gloucester, UK: Sutton Publishing.
founded on 1002 consecutive cases. J Hepatol 3:154 163. Roberts C, Lucy D, Manchester K. 1994. Inammatory lesions of
Mani S, McDermott S, Valtorta M. 1997. MENTOR: a Bayesian the ribs: an analysis of the Terry collection. Am J Phys An-
model for prediction of mental retardation in newborns. Res thropol 95:169 182.
Dev Disabil 18:303318. Rogers J, Waldron T, Dieppe P, Watt I. 1987. Arthropathies in
McNeil BJ, Sherman H. 1978. Example: Bayesian calculations for palaeopathology: the basis of classication according to most
the determination of the etiology of pleuritic chest pain in probable cause. J Archaeol Sci 14:179 193.
young adults in a teaching hospital. Part B. Comput Biomed Rothschild BM, Woods RJ, Ortel W. 1990. Rheumatoid arthritis
Res 11:187194. in the buff: erosive arthritis in deeshed bones. Am J Phys
Miettinen OS, Caro JJ. 1994. Foundations of medical diagnosis: Anthropol 82:441 450.
what actually are the parameters involved in Bayes theorem. Russek E, Kronmal RA, Fisher LD. 1983. The effect of assuming
Stat Med 13:201209. independence in applying Bayes theorem to risk estimation and
Miller E, Ragsdale BD, Ortner DJ. 1996. Accuracy in dry bone classication in diagnosis. Comput Biomed Res 16:537552.
diagnosis: a comment on paleopathological methods. Int J Os- Schell LM, Ulijaszek SJ, editors. 1999. Urbanism, health and
teoarchaeol 6:221229. human biology in industrialized countries. Cambridge: Cam-
bridge University Press.
Montironi R, Bartels PH, Thompson D, Scarpelli M, Hamilton
Sebes JI, Diggs LW. 1979. Radiographic changes in the skull in
PW. 1994. Prostatic intraepithelial neoplasia. Development of a
sickle-cell anemia. AJR 132:373377.
Bayesian belief network for diagnosis and grading. Anal Quant
Sherman H. 1978. A pocket diagnostic calculator program for
Cytol Histol 16:101112.
computing Bayesian probabilities for nine diseases with six-
Montironi R, Mazzucchelli R, Santinelli A, Hamilton PW, Thompson
teen symptoms. Comput Biomed Res 11:177186.
D, Batels PH. 1998. Case diagnosis as positive identication in
Siles S, Garrigues V, Ponce J, Galvez C, Berenguer J. 1997. Analysis
prostatic neoplasia. Anal Quant Cytol Histol 20:424 436.
of the predictive value of clinical data in patients with suspected
Norusis MJ, Jacquez JA. 1975. Diagnosis. I. Symptom noninde-
colonic disease. Rev Esp Enferm Apar Dig 89:445 456.
pendence in mathematical models for diagnosis. Comput Somogyi L, Cikes N, Marusic M. 1993. Evaluation of criteria
Biomed Res 8:156 172. contributions for the classication of systemic lupus erythem-
Nugent CA, Warner HR, Dunn JT, Tyler FH. 1964. Probability atosus. Scand J Rheumatol 22:58 62.
theory in diagnosis of Cushings Syndrome. J Clin Endorcrinol Starmer CF, Lee KL. 1976. A mathematical approach to medical
Metab 24:621 623. decisions: application of Bayes rule to a mixture of continuous
Ohmann C, Yang Q, Kunneke M, Stoltzing H, Thon K, Lorenz W. and discrete clinical variables. Comput Biomed Res 9:531541.
1988. Bayes theorem and conditional dependence of symptoms: Steinbock RT. 1976. Paleopathological diagnosis and interpreta-
different models applied to data of upper gastrointestinal tion. Springeld, IL: C.C. Thomas.
bleeding. Methods Inf Med 27:73 83. Van de Merwe JP. 1983. Differentiation between two diseases
Ortner DJ. 1991. Theoretical and methodological issues in paleo- using a programmable hand-held calculator and Bayes theo-
pathology. In: Ortner DJ, Aufderheide AC, editors. Human rem: application to Crohns disease and ulcerative colitis. Com-
paleopathology: current syntheses and future options. Wash- put Biol Med 13:317332.
ington, DC: Smithsonian Institution Press. p 511. Vishnevskiy AA, Artobolevskiy II, Bykowskiy ML. 1973. Princi-
Ortner DJ. 1992. Skeletal paleopathology: probabilities, possibil- ples of a solution to the problem of diagnosis using digital
ities, impossibilities. In: Verano JW, Ubelaker DH, editors. computers. In: AA Vishnevskiy, editor. Machine diagnosis and
Disease and demography in the Americas. Washington, DC: information retrieval in medicine in the USSR. Bethesda, MD:
Smithsonian Institution Press. p 513. U.S. Department of Health, Education, and Welfare. p 113.
Ortner DJ. 1994. Descriptive methodology in paleopathology. In: Waldron T. 1994. Counting the dead. The epidemiology of skeletal
Owsley DW, Jantz RL, editors. Skeletal biology in the Great populations. Chichester, UK: John Wiley & Sons.
Plains. Washington, DC: Smithsonian Institution Press. p 73 80. Warner HR, Totonto AF, Veasey LG, Stephenson R. 1961. A math-
Ortner DJ, Putschar WGJ. 1981. Identication of pathological ematical approach to medical diagnosis. JAMAf 177:177183.
conditions in human skeletal remains. Smithsonian contribu- Warner HR, Totonto AF, Veasey LG. 1964. Experience with
tions to anthropology, no. 28. Washington, DC: Smithsonian Bayes theorem for computer diagnosis of congenital heart dis-
Institution Press. ease. Ann NY Acad Sci 115:558 567.
Overall JE, Williams CM. 1963. Conditional probability program Whipple GH, Bradford WL. 1932. Racial or familial anemia of
for diagnosis of thyroid function. JAMA 183:307. children. Am J Phys Anthropol 44:336 365.
Pastor A, Menendez R, Cremades MJ, Pastor V, Llopis R, Aznar Worbel JS, Connolly JE. 1998. Making the diagnosis of osteomy-
J. 1997. Diagnostic value of SCC, CEA and CYFRA 21.1 in lung elitis. The role of prevalence. J Am Podiatr Med Assoc 88:337
cancer: a Bayesian analysis. Eur Respir J 10:603 609. 343.
Perpina M, Pellicer C, de Diego A, Compte L, Macian V. 1993. Yan YY. 2000. The inuence of weather on human mortality in
Diagnostic value of the bronchial provocation test with metha- Hong Kong. Soc Sci Med 50:419 427.
choline in asthma. A Bayesian analysis approach. Chest 104: Zimmerman MR, Kelley MA. 1982. Atlas of human paleopathol-
149 154. ogy. New York: Praeger.

Vous aimerez peut-être aussi