Weincludedinouranalysispreoperativedatasuchasdemographiccharacteristics,laboratoryfindings,comorbidities,andmedication,aswellasperioperativedatasuchasdurationofsurgery.Thebaselinecharacteristicsincludedage,gender,anddyslipidemia.Dataontumorcharacteristicssuchasalpha-fetoprotein(AFP)andtumorsizewerealsocollected.Laboratorymeasurementsincludedhemoglobin,serumcreatinine,andcholesterol.Perioperativevariablesincludedtheuseofbloodproductsandsurgeryduration.
Serumcreatininefluctuationswerecontinuouslymonitoredaftertheoperation,andwerecomparedwiththepreoperativebaselinevalues.OurresultsindicatethataLSA-AKIeventoccurredin296patients(12.1%)within7daysaftersurgery.TheincidenceofAKIinthetrainingsetandtestsetwas11.5%(198/1715)and13.3%(98/735),respectively.
Therearealsosomelimitationsinourstudy.First,thiswasasingle-centerretrospectivestudy.Duetotherelativelysmallsamplesizeandthelackofexternalvalidation,ourresultsmaynotbegeneralizable.Second,includingallvariablesintheprocessofdatacollectionisaverychallengingtask,andthereforesomepotentiallyrelevantfactorsmayhavebeenignored.Finally,mostoftheinputtedfeatureswereimplementedmanually.Wearestillworkingondevelopingareal-timeautomatedelectronichealthrecordalgorithmthatcouldcollectperioperativepatientinformationfromavarietyofdatasources.Withthesenewtechnologies,predictivemodelsbasedonmachinelearningmayhavethepotentialtochangeclinicalpractice.
Recently,machinelearninghasprovenhelpfulintheinterpretationofmedicalresultsandhaspotentialforhelpingguidediagnosisandtreatment,ultimatelyimprovingpatientoutcomes.
Machinelearningmethodstopredictacutekidneyinjury(AKI)eventsremainlargelyunexplored.
WeaimedtodeveloppredictionmodelsforAKIafterlivercancerresectionbasedonmachinelearningtechniques.
Atotalof2450patientswhohadundergoneprimaryhepatocellularcarcinomaresectionatChangzhengHospital,ShanghaiCity,China,fromJanuary1,2015toAugust31,2020werescreened.Patientswererandomlyassignedtothetrainingandthetestsetsataratioof7:3.Thetrainingsetwasusedformodeldevelopmentandoptimization,whilethetestsetwasusedformodelvalidationandevaluation.
AKIeventsoccurredin296patients(12.1%)aftersurgery.Amongtheoriginalmodelsbasedonmachinelearningtechniques,therandomforest(RF)algorithmhadoptimaldiscriminationwithanareaunderthecurvevalueof0.92,comparedto0.87forextremegradientboosting,0.90fordecisiontree,0.90forsupportvectormachine,and0.85forlogisticregression.TheRFalgorithmalsohadthehighestconcordance-index(0.86)andthelowestBrierscore(0.076).ThevariablesthatcontributedthemostintheRFalgorithmwereage,cholesterol,andsurgerytime.
MachinelearningtechnologycanaccuratelypredictAKIafterhepatectomy.
Intheeraofpersonalizedmedicine,ourmodelbasedonmachinelearningcandiscriminatepatientsathighriskforAKI,thushelpingguideclinicaldecisionsandfacilitatingprospectiveinterventionsforhigh-riskindividuals.
Provenanceandpeerreview:Unsolicitedarticle;Externallypeerreviewed.
Peer-reviewmodel:Singleblind
Specialtytype:Urologyandnephrology
Country/Territoryoforigin:China
Peer-reviewreport’sscientificqualityclassification
GradeA(Excellent):0
GradeB(Verygood):0
GradeC(Good):C
GradeD(Fair):0
GradeE(Poor):0
P-Reviewer:VeelkenRS-Editor:GongZML-Editor:WebsterJRP-Editor:GongZM