Science - USA (2022-02-25)

SCIENCE science.org

GRAPHIC: V.ALTOUNIAN/

SCIENCE

pen when drugs are prescribed to provide

rapid clinical intervention while await-

ing results from susceptibility tests ( 4 ).

However, a substantial fraction of cases fall

into the insidious category of being appro-

priately treated with susceptibility-matched

antibiotics and yet returning with an antibi-

otic-resistant infection.

How do these resistant strains emerge?

Stracy et al. carried out genomic sequenc-

ing of the bacteria, providing a detailed

view of the strains and species of the origi-

nal infection compared with the ones that

caused it to recur. This analysis reveals an

underappreciated path to reinfection, with

the original species being treated and elimi-

nated but with the treatment ultimately

setting the stage for other resistant strains

to emerge. Recent studies provide compel-

ling examples of how a patient’s microbiota

serves as a reservoir, such as uropathogenic

strains being harbored in the gut ( 2 , 5 ). This

analysis also joins a growing list of studies

that highlight the power and potential of

integrating genome sequencing data with

clinical records ( 6 , 7 ).

Building on the extensive dataset of in-

fection and treatment histories, Stracy et al.

develop a machine learning algorithm for

patient-specific recommendations (see the

figure). This strategy is successful because

the patient’s risk of developing a recurrent

infection is strongly linked to their treat-

ment and infection history. The authors use

multivariate logistic regression, a powerful

long-established tool for medical outcome

predictions ( 8 ), to assess the risk of early re-

currence for each antibiotic candidate. The

algorithm then identifies the antibiotic that

is least likely to cause an infection to recur

early and be resistant to the prescribed

antibiotic. Notably, the model uses only

13 parameters to fit the data for each UTI

antibiotic, and eight in the case of wound

infections, in which the parameters weight

factors such as age, sex, pregnancy status,

catheter use, and infection history.

These recommendation algorithms are

evaluated against statistical models of early

recurrence and are predicted to reduce the

rate of early resistance recurrences of UTIs

and wound infections by almost half com-

pared with those of physicians’ suscepti-

bility-matched prescriptions. This is one

of the great advantages of machine learn-

ing approaches: By systematically review-

ing all available patient information, with

a statistical model built on thousands of

training samples, the algorithm can iden-

tify subtle patterns and provide valuable

insights and recommendations to support

physicians in their decisions ( 4 , 9 ). The al-

gorithms are also predicted to significantly

improve on current physician practices

even when they are constrained to recom-

mend different types of antibiotics at fre-

quencies that match present-day practices.

This is essential because other factors,

such as side effects and ease of use, inform

decisions about antibiotic selection. The

risk assessment for each antibiotic and a

quantified contribution profile for each pa-

tient risk factor, such as age or resistance

history, can be retrieved. The decision is

therefore interpretable, a critical step to-

ward establishing trust in machine learn-

ing recommendation systems for health

care ( 10 , 11 ).

However, these algorithms are not ready

to prescribe antibiotics just yet. The al-

gorithms were not evaluated on actual

patients but against statistical estimates

of the chances of early resistance recur-

rences. Further studies are required to as-

sess whether resistant infections that ex-

tend beyond the 28-day window are also

prevented. It is also unknown how large

and diverse the training sets need to be

and, once trained, how robust the recom-

mendation system is, for example, to miss-

ing data points or clerical errors.

More generally, data-driven approaches

to health care decisions face numerous

social, legal, and ethical challenges, as

highlighted recently by the World Health

Organization (WHO) ( 12 ). Algorithms can

entrench suboptimal or even discrimina-

tory practices ( 13 ). Although the recom-

mendation model presented by Stracy et

al. is not fitted directly to the clinicians’

decisions if, for example, the responses of

10- to 19-year old women to fosfomycin are

underrepresented in the training set, pre-

dictions under these conditions will be less

reliable. It is also unclear how the model

will adapt to shifts in medical practices—

for example, in the case of newly available

antibiotics. Still, machine learning recom-

mendation systems such as the one pre-

sented by Stracy et al. have the potential

to substantially improve patient outcomes

and could play a major role in mitigating

antibiotic resistance.

The current wave of data-driven meth-

ods for health care presents an opportu-

nity not only to tackle broad public health

issues but to harness their power to intro-

duce new data collection standards, evalu-

ation metrics, and training paradigms

( 14 , 15 ). Essential to the success of these

methods is the availability of both lon-

gitudinal datasets and interdisciplinary

studies, such as those based on genomic

sequence analysis, that provide fundamen-

tal insights into the mechanisms by which

resistance emerges. j

REFERENCES AND NOTES

1. A. L. Flores-Mireles, J. N. Walker, M. Caparon, S. J.
   Hultgren, Nat. Rev. Microbiol. 13 , 269 (2015).


2. B. M. Forde et al., Nat. Commun. 10 , 3643 (2019).


3. M. Stracy et al., Science 375 , 889 (2022).


4. I. Yelin et al., Nat. Med. 25 , 1143 (2019).


5. K. L. Nielsen, P. Dynesen, P. Larsen, N. Frimodt-Møller, J.
   Med. Microbiol. 63 , 582 (2014).


6. X. Didelot, R. Bowden, D. J. Wilson, T. E. A. Peto, D. W.
   Crook, Nat. Rev. Genet. 13 , 601 (2012).


7. M. Magruder et al., Nat. Commun. 10 , 5521 (2019).


8. P. W. F. Wilson et al., Circulation 97 , 1837 (1998).


9. A. Rajkomar, J. Dean, I. Kohane, N. Engl. J. Med. 380 , 1347
   (2019).


10. C. M. Cutillo et al., NPJ Digit. Med. 3 , 47 (2020).


11. A. F. Markus, J. A. Kors, P. R. Rijnbeek, J. Biomed. Inform.
    113 , 103655 (2021).


12. WHO, Ethics and governance of artificial intelligence
    for health: WHO guidance (WHO, 2021); http://www.who.int/
    publications/i/item/9789240029200.


13. Z. Obermeyer, B. Powers, C. Vogeli, S. Mullainathan,
    Science 366 , 447 (2019).


14. J. F. Rajotte et al., GoodIT 2021 Proc. 2021 Conf. Inf.
    Technol. Soc. Good, 10.1145/3462203.3475875 (2021).


15. S. Hooker, Patterns 2 , 100241 (2021).

10.1126/science.a bn9969

Past infections Current infection

Age, sex,

catheter use,

pregnancy

status

Machine learning model

Antibiotic recommendation for current infection

Antibiotic

susceptibilities

A

B

C

D

E

...

A

B

C

D

E

...

A

Susceptible ResistantRRRRReRe Prescribed antibiotic

0 1

Probability of resistance

Predicting resistance according

to treatment history

A machine learning model built with data from

thousands of patients with wound infections and

urinary tract infections identified the factors

that contribute to antibiotic resistance in recurrent

infections. A patient’s history of infections and

antibiotic treatments can then be used with their

demographic data to predict which candidate

antibiotics would likely prevent a recurrent infection.

25 FEBRUARY 2022 • VOL 375 ISSUE 6583 819