Question

A logistic regression is a composed of a uniquely identifying number, followed by multiple binary variables (always 1 or 0) based on whether or not a person meets certain criteria. Below I have a query that lists several of these binary conditions. With only four such criteria the query takes a little longer to run than what I would think. Is there a more efficient approach than below? Note. tblicd is a large table lookup table with text representations of 15k+ rows. The query makes no real sense, just a proof of concept. I have the proper indexes on my composite keys.

select  patient.patientid 
,case when exists
(
    select c.patientid from tblclaims as c
    inner join patient as p on p.patientid=c.patientid
    and c.admissiondate = p.admissiondate
    and c.dischargedate = p.dischargedate
    where patient.patientid = p.patientid
    group by c.patientid
    having count(*) > 1000
    )
    then '1' else '0'
    end as moreThan1000
,case when exists
(
    select c.patientid from tblclaims as c
    inner join patient as p on p.patientid=c.patientid
    and c.admissiondate = p.admissiondate
    and c.dischargedate = p.dischargedate
    where patient.patientid = p.patientid
    group by c.patientid
    having count(*) > 1500
    )
    then '1' else '0'
    end as moreThan1500
,case when exists
(
    select distinct picd.patientid from patienticd as picd
    inner join patient as p on p.patientid= picd.patientid
    and picd.admissiondate = p.admissiondate
    and picd.dischargedate = p.dischargedate
    inner join tblicd as t on t.icd_id = picd.icd_id
    where t.descrip like '%diabetes%' and patient.patientid = picd.patientid
    )
    then '1' else '0'
    end as diabetes
,case when exists
(
    select r.patientid, count(*) from patient as r
    where r.patientid = patient.patientid
    group by r.patientid
    having count(*) >1
    ) 
    then '1' else '0'
    end 


from patient
order by moreThan1000 desc

Answer 1

I would start by using subqueries in the from clause:

select q.patientid, moreThan1000, moreThan1500,
       (case when d.patientid is not null then 1 else 0 end),
       (case when pc.patientid is not null then 1 else 0 end)
from patient p left outer join
     (select c.patientid,
             (case when count(*) > 1000 then 1 else 0 end) as moreThan1000,
             (case when count(*) > 1500 then 1 else 0 end) as moreThan1500
      from tblclaims as c inner join
           patient as p
           on p.patientid=c.patientid and
              c.admissiondate = p.admissiondate and
              c.dischargedate = p.dischargedate
      group by c.patientid
     ) q
     on p.patientid = q.patientid left outer join
     (select distinct picd.patientid
      from patienticd as picd inner join
           patient as p
           on p.patientid= picd.patientid and
              picd.admissiondate = p.admissiondate and
              picd.dischargedate = p.dischargedate inner join
          tblicd as t
          on t.icd_id = picd.icd_id
      where t.descrip like '%diabetes%'
     ) d
     on p.patientid = d.patientid left outer join
     (select r.patientid, count(*) as cnt
      from patient as r
      group by r.patientid
      having count(*) >1
     ) pc
     on p.patientid = pc.patientid
order by 2 desc

You can then probably simplify these subqueries more by combining them (for instance “p” and “pc” on the outer query can be combined into one). However, without the correlated subqueries, SQL Server should find it easier to optimize the queries.